To date, the gold standard method for assessing attachment in infants is the Strange Situation Procedure (SSP) using the Ainsworth scales for coding organised attachment (Ainsworth et al., 1978) and the Main and Solomon (1990) indices for disorganisation and disorientation. Attachment behaviours in the preschool period present differently from those in infancy, necessitating assessment through a modified method (Cassidy & Marvin, 1992) . At four years of age, language capabilities, perspective-taking abilities and goal-corrected behaviours are well developed, together with some degree of self-regulation (Conti-Ramsden & Durkin, 2012; Marvin, 1977) . Language is often used by the preschooler in place of physical proximity-seeking to re-establish psychological connection with their caregiver following separation. Avoidance is more subtle and nuanced in the preschooler, reflecting learned representations of social norms (e.g., that ignoring an adult is socially inappropriate). By four years of age, disorganised attachment behaviours predominantly shift to controlling forms of behaviour, maintaining a lack of coherence characteristic of infant attachment disorganisation. These changes add significant complexity to the assessment of attachment in the preschool years. As a result, the gold-standard Ainsworth infant coding system and the Main and Solomon (1990) disorganised index no longer apply for preschoolers.

Three validated methods of classifying preschool attachment have predominated the literature to date: the Attachment Q-Sort (Waters & Deane, 1985) , the Preschool Assessment of Attachment (Crittenden, 1992) , and the Preschool Attachment Classification System (Cassidy & Marvin, 1992) . Of these, the PACS remains the gold-standard assessment for preschool attachment classification, and is utilized by trained and reliable coders, based on a modified SSP. For each, an attachment classification is determined: Secure (B), Avoidant (A), Ambivalent/Resistant (C), Disorganised (D), and Insecure/Other (I/O).

The coding of attachment behaviours using the PACS involves careful attention to behavioural markers such as proximity and contact, body orientation, verbal conversation/discourse, gaze, and affect. Coders note such behavioural markers demonstrated by the child during the SSP toward their attachment figure with specific attention to the two reunion episodes, then compare their formulations to qualitative descriptions of each attachment classification (and sub-classification) in the PACS manual (Cassidy & Marvin, 1992) .

A set of continuous rating scales based on the PACS was proposed by Moss and colleagues (2015) in the Preschool Attachment Rating Scales (PARS). The PARS was found to be a reliable and valid tool for the assessment of preschool attachment, demonstrating good convergent validity with independently coded categorical attachment (Deneault et al., 2020) . However, as with the PACS, the PARS is a similarly challenging and resource-intensive method to learn and employ, and founded on USA data.

Use of formal attachment assessment methods is predominantly confined to developmental research settings. Obstacles are many, and include the time and expense of obtaining reliability in the PACS. New coders are typically required to undertake a two-week training program with a certified trainer, followed by successful completion of multiple preschool attachment cases, involving at minimum 80% match to the trainer’s codes. Coding workshops are not frequently offered and involve travel and extended time away. Unlike the infant attachment coding system (Ainsworth et al., 1978) , preschool coders must distinguish qualitative differences between attachment groups, increasing the likelihood of misclassification. Finally, the validity of scales such as the PARS has only been assessed among reliable coders. Their validity for clinician use remains untested, and fit for purpose with Australian data is unestablished. To address these limitations, the APAS (McIntosh et al., 2017) was developed to 1) facilitate a large research task in the Australian Temperament Project of classifying preschool attachment organisation in our Generation 3 study (see Method for details), and 2) assist trainee clinicians in our lab, in the accurate observation of relevant attachment behavioural patterns. Below we report on first validation data of the tool for these purposes.

The specific aims of the study were threefold: 1) to examine the reliability and validity of the APAS compared to the PACS when both were assessed by certified coders; 2) to compare attachment classification results on the APAS between certified coders of preschool attachment and clinical psychologists previously naïve to preschool attachment classification, and 3) to discuss the implications of findings for enhancing clinical assessment of attachment security in the preschool years.

Participants were 121 four-year-old child-parent dyads drawn from a nested observational study of attachment, completed at four years of age, within a larger Australian intergenerational study of social and emotional development, The Australian Temperament Project Generation 3 Study. The cohort commenced in 1983 with recruitment through geographically distributed maternal and child health centres in the state of Victoria. Generation 1 (G1) parents were approached for participation four to eight months after the birth of the study child (Generation 2: G2). The study has since tracked the social-emotional development of the main G2 cohort across 38 years. The initial cohort comprised 2443 G2 infants from urban and rural areas of Victoria, Australia. Details on recruitment strategy, sampling and sample characteristics are provided elsewhere (Prior, Sanson, Smart, & Oberklaid 2000) . Attrition in the cohort has been approximately 1% per annum (Letcher et al., 2012) .

Since 2011, over 1100 Generation 3 (G3) offspring born to G2 participants have been identified through a systematic screening process and assessed in pregnancy, at 10 weeks postpartum, and at 1 and 4 years of age. A nested sample within the Generation 3 cohort participated in two observations of attachment and caregiving when the G3 child was one and four years of age (Life@1 and Life@4 assessments respectively). The current study comprises all cases available during the study window, namely the first 121 G2/G3 child-parent dyads who participated in the Life@4 observational study of attachment. Specifically, 98 families including 23 couples (32 fathers and 89 mothers) participated with a total of 109 children (48 boys and 61 girls; 57% first born), 12 children participated with both father and mother. At the time of the SSP assessment, the average age of children was 4.31 years (SD = 0.18). All parents in this study were born in Australia. Of the 121 parents, 64% were from born from two Australian parents, and 36% had mixed ancestry.

G2 participants in the ATP-G3 study were invited to a play session within two months of their G3 child’s fourth birthday (Life@4). Filmed observations of child-parent dyads were conducted using the Cassidy and Marvin (1992) modified SSP for preschoolers. Each Life@4 SSP video recording was analyzed by staff in the ATP-G3 Melbourne Attachment and Caregiving (MAC) Lab and coded using the purpose-designed APAS coding sheet. Staff comprised two groups: reliable coders trained by Professor Emeritus Robert Marvin and certified in the PACS (Cassidy & Marvin, 1992) , and naïve coders who were clinical psychology postgraduate students, trained in the general concepts of attachment theory and observations, but not certified in observational attachment assessment. APAS cases included in this pilot study were independently coded by naïve and reliable coders (the latter including authors, ET, AB, JO and JM). All coders viewed the full SSP recording at least once before assigning scores on the APAS. Emphasis for scoring was placed on the two reunion episodes, although indicators of disorganisation or disorientation were coded based on any instance during the SSP where the child was with their parent. Following completion of the APAS, reliable coders assigned an attachment classification (and sub-classification) based on the Cassidy and Marvin (1992) PACS.

Of the 121 child-parent dyads, 38.8% were classified Secure (n = 47), 26.5% Avoidant (n = 32), 13.2% Ambivalent-Resistant (n = 16), and 21.5% Disorganised/Insecure-Other (n = 26) on the PACS. Attachment classification distributions did not differ for mother-child and father-child dyads, χ² (3) = 3.24, p = 0.524, or by child gender χ² (3) = 2.95, p = 0.399.

Table 1 presents the Cronbach’s alpha and inter-item correlation coefficients of the five BAC Subscales (Proximity and Contact, Body Orientation, Speech, Gaze, and Affect) and the BAC Total scale for reliable and naïve coders. All Cronbach’s alpha values were acceptable (i.e., α > 0.70 as recommended by McCrae et al., 2011) . Inter-item correlation coefficients were largely within the acceptable range (i.e., r = 0.15 - 0.50 as recommended by Clark & Watson, 1995 ). The Body Orientation subscale was just out of this range (r = 0.52 for both reliable and naïve coders). The Gaze subscale was above the recommended range (r = 0.74 for reliable coders, and 0.70 for naïve coders), potentially indicating item redundancy within this subscale.

Table 2 presents the Pearson correlation coefficients comparing reliable and naïve coders’ scores for each APAS BAC subscale in each attachment classification (i.e., Secure, Avoidant, and Ambivalent-Resistant). BAC scores were highly correlated (r = 0.52 - 0.79), with the exception of the Gaze subscale for Ambivalent-Resistant attachment at .18. This indicates strong alignment between reliable and naïve coders.

Figure 1 presents the results of the linear regression analyses illustrating differences

Note. p < 0.05 for all correlations reported.

between BAC scores (determined by non-overlapping confidence intervals) across the formal PACS attachment classifications. The mean BAC score refers to the correlation with the proto-typical item scores given by reliable coders to each of the Secure, Avoidant, and Ambivalent-Resistant categories. The corresponding 95% confidence intervals were plotted for the five BAC subscales and the BAC Total scale.

We found that all BAC subscale and Total scores with the Secure attachment prototype were highest in children classified as Secure by the PACS method. Similarly, all BAC subscales and Total scores with the Avoidant attachment prototype were highest for children classified Avoidant in the PACS method. Patterns were slightly less consistent for BAC scores with the Ambivalent-Resistant attachment prototype. Specifically, three of the five subscales (i.e., Proximity and Contact, Speech, Affect) and Total scores with the Ambivalent-Resistant attachment prototype were highest for children with ambivalent-resistant attachment. The Body Orientation and Gaze subscale scores did not clearly distinguish children with ambivalent-resistant attachment from children with secure or avoidant attachments. Overall, results were consistent across both reliable and naïve coders.

The Youden index from the ROC curve analysis indicated that a score of 4 or above distinguished Organised from Disorganised and Insecure-Other (D/IO) attachments (AUC = 0.81 [95% CI 0.71 - 0.90], sensitivity = 0.72, specificity = 0.74).

The organised attachment classifications assigned from the APAS rating scales by reliable and naïve coders are presented in Table 3 (i.e., excluding D/IO attachment classifications), and in Table 4 for all attachment classifications (i.e.,

including D/IO attachment classifications). Each table presents findings separately for reliable coders and naïve coders on the APAS, for accuracy of the APAS derived classification with that of the reliable coder’s derived classification in the PACS method.

Using Cohen’s suggestions for interpretation of the Kappa statistic (Cohen, 1960; McHugh, 2012) , the assignment of organised attachment classifications from the APAS rating scales by reliable coders had almost perfect accuracy: κ = 0.89 with 93.68% agreement. Naïve coders had high accuracy in assigning the correct attachment: κ = 0.71 with 83.16% agreement. Accuracy was lower when

incorporating D/IO attachment classifications. Accuracy was in the moderate-to-substantial range for reliable coders (κ = 0.61 with 72.73% agreement), and in the low-to-moderate range for naïve coders (κ = 0.46 with 62.81% agreement).

The five BAC subscales (i.e., Proximity and Contact, Body Orientation, Speech, Gaze, and Affect) demonstrated strong internal consistency. The inter-item correlation coefficients for the five BAC subscales were within acceptable range, except for the Body Orientation and Gaze subscales. This suggests item redundancy within these subscales. Further item refinement using a larger sample may be possible, to accurately capture the distinguishing behaviours.

The APAS demonstrated sound discriminant ability, with all coders (with or without certification in the formal classification system) accurately distinguishing the patterns of behaviours associated with secure, avoidant and ambivalent/resistant attachment classifications. Findings suggest that the APAS BAC subscales are particularly useful in identifying attachment behaviours relevant for secure and avoidant classifications. Our results show that accurate identification of ambivalent-resistant attachments rests heavily on the Proximity, Speech and Affect subscales. Increased weighting on these three BAC subscales may be warranted, given lower differentiation on the Body Orientation and Gaze subscales.

This finding aligns with the classification dilemmas faced using other coding methods, given children with ambivalent-resistant attachments prototypically demonstrate a mix of approach and avoidance/resistance to their caregiver when in need. Scoring these conflicting behaviours may prove challenging, even for reliable coders, and attention to the patterning of behaviours across all subscales is ultimately needed in the PACS as well as the APAS.

Using the APAS “BAC Total score”, findings showed accurate assignment of organised three-way attachment classifications for both reliable and naïve coders. Combining the APAS BAC and D/IO Index, the accuracy of assignment of disorganised attachment classifications remained in the moderate range, but was slightly lower. This may be due to the piloted scoring approach of the D/IO Index, where each item was assigned a binary score. We have since modified this approach to a 5-point scale which provides more nuanced ratings of each item on the D/IO Index.

The study has a number of strengths. First, the APAS is theoretically and empirically based on the current gold-standard observational assessment of preschool attachment, the Cassidy & Marvin (1992) PACS. Second, the independent completion of the APAS by reliable and naïve coders allowed for comparison of scores and strengthens the potential clinical utility of this new instrument. Third, the naïve coders were postgraduate students in training to be clinicians. The findings from our study therefore suggest the potential of the APAS to refine clinicians’ assessment of preschool attachment. In the first place, becoming familiar with the items and sub-scales in Supplementary Figure S1 and Figure S2 may enhance the clinician’s focus on relevant behaviours within an observation of child-parent interaction. Care must be taken to establish the observational conditions under which attachment behaviours are elicited, as per the conditions of the pre-school SSP.

However, there are important limitations to note. All items are behaviourally specific, rated for the extent to which they were present. The inclusion of items aiding appraisal of the coherence and timing of certain behaviours may assist identification of Ambivalent-Resistant and Disorganised attachment groups. Our naïve coders were all clinical psychology postgraduate students and our findings may not apply to other fields of training. We emphasise that clinical reasoning remains central in the classification of complex attachment presentations, and this is predicated on clinical training and expertise in socio-emotional development. Until replication studies are completed in outer social science and therapeutic disciplines, caution is warranted in the use of the APAS. Further, assigning formal classifications for any legal or statutory purpose through the APAS alone is not supported. The APAS was designed to enable better recognition of key attachment behaviours; use for clinical diagnosis of attachment-related disorders or neurodevelopmental disorders is neither conceptually nor empirically supported at this time.