Exploring the meaning of unresolved loss and trauma in more than 1,000 Adult Attachment Interviews

Introduction
Unresolved states of mind about loss/abuse (U/d) are identified through lapses in the monitoring of reasoning, discourse, and behaviour surrounding loss/abuse in response to the Adult Attachment Interview (AAI; George et al.,1996; Main et al., 2003). Many studies have established that the unresolved category is predictive of infant disorganised attachment (Van IJzendoorn et al., 1995; Verhage et al., 2016). However, relevant research is based upon the assumption that unresolved states of mind are a unitary construct underlying a large variety of indicators. Although the coding system for assessing unresolved loss/abuse has been widely used for decades, the individual indicators have not been validated independent of the development sample. This study used statistical (machine) learning techniques to explore potential associations between indicators of unresolved loss/abuse, unresolved states of mind, and infant disorganised attachment. The study protocol was preregistered on Open Science Framework.

Research questions
1. To what extent are different indicators of unresolved loss/abuse associated with scores and classifications of unresolved states of mind?
2. To what extent do indicators of unresolved loss/abuse predict infant disorganised attachment classifications?

Study population
AAI data were included from 1,009 caregiver-child dyads from 13 studies in the Collaboration on Attachment Transmission Synthesis (Verhage et al., 2018). Infant disorganised attachment data (measured in the Strange Situation, Ainsworth et al., 1978) was available from 930 dyads.

Methods
Statistical learning models were developed to predict unresolved score, based on the following dichotomous variables: indicators of unresolved loss/abuse (lapse present/not present in interview), risk status (normative/high-risk), reports of loss or abuse (loss or abuse reported/not reported), and study sample (dummy-coded). Models were estimated using a 70% training set and tested on unseen data (30% testing set). The predicted unresolved scores were then used to predict unresolved classification and infant disorganised attachment, aiming to achieve high sensitivity and specificity for prediction of the classifications.

Results
Twenty-three percent of adults were classified as unresolved, and 23% of infants were classified as disorganised. Because of near-zero occurrence, only 10 of the 16 indicators of unresolved loss and 2 of the 16 indicators of unresolved abuse were included in the predictive models. The best predictive model was the lasso model (a method for penalised regression). All included indicators of unresolved loss/abuse contributed to predicting actual unresolved score (Table 1). The lasso model correctly identified 41% of actual unresolved classifications at a threshold predicted unresolved score of ≥ 5 (the AAI coding manual’s threshold for classification of unresolved), meaning that 59% of unresolved cases were missed. Nearly all (98%) not-unresolved cases were correctly identified by the model. The lasso model correctly predicted 20% of actual infant disorganised attachment classifications and 85% of actual not-disorganised classifications. Figure 1 shows the trade-off between sensitivity and specificity for infant disorganised classification, based on participants’ predicted unresolved scores using the indicators of unresolved loss/abuse.

Conclusions
Results will be discussed with regard to the contribution of individual indicators in shaping the construct of U/d and overfitting of the AAI coding system to Main and colleagues’ (1985) development sample.