What the Research Measures
Attachment research operates across two largely separate methodological traditions that have different strengths and ask somewhat different questions. Understanding which tradition a finding comes from matters for interpreting it.
The interview tradition uses the Adult Attachment Interview (AAI) — a semi-structured clinical interview scored for narrative coherence rather than content. It assesses adults' mental representations of their own childhood attachment experiences and is the gold standard for clinical and developmental research. The AAI predicts parenting behavior, infant Strange Situation classification, and therapeutic outcomes, but it requires trained coders and is impractical for large-scale studies.
The self-report tradition uses instruments like the ECR-R (Experiences in Close Relationships — Revised) and its predecessors. These assess current feelings and patterns in romantic relationships rather than representations of the past. They are practical for large studies and have excellent psychometric properties — the ECR-R shows Cronbach's alphas typically above .90 for both subscales — but they capture something different from the AAI. The cross-classification agreement between AAI categories and ECR-R quadrant placement is lower than researchers initially hoped, which has generated ongoing debate about whether the two traditions are measuring the same underlying construct or adjacent but distinct phenomena.
Most of what readers encounter in popular attachment content — the four-style typology, the quiz-based assessments, the relationship research — comes from the self-report tradition. The clinical research on therapy outcomes and parenting transmission comes primarily from the interview tradition. This page draws on both.
Stability and Change
One of the most practically important questions in attachment research — and one of the most technically contested — is how stable attachment patterns are over time and whether they can change meaningfully in adulthood.
The short answer from the literature: attachment is moderately stable, more stable in adulthood than in childhood, but not fixed. Change is real and documented, driven primarily by significant relationship experiences and explicit reflective work.
The more technically precise answer comes from Fraley's (2002) meta-analysis and subsequent work. Fraley integrated longitudinal data from 27 effect sizes and found that the test-retest stability of secure vs. insecure classification from infancy through adulthood showed a weighted mean correlation of approximately r=.39 — substantial consistency, but nowhere near the stability of basic personality traits like Big Five Conscientiousness, which typically shows test-retest correlations above .70 over comparable intervals. Crucially, Fraley found that stability correlations tended to plateau rather than continue declining as the test-retest interval extended — a pattern more consistent with what he called the prototype model: a stable underlying attachment representation with variation fluctuating around it, rather than a system that either stays fixed or gradually revises based on new experience.
In adulthood specifically, Fraley and colleagues found evidence for what they termed "stable instability" — attachment scores that fluctuate meaningfully around a stable individual baseline. This means short-term variations in how secure or anxious someone feels in their relationship are real, driven by day-to-day relational experiences, but they tend to return toward a personal set point. The set point itself can shift, but the shift requires sustained experience, not isolated events.
What drives change? The evidence points to three main routes: significant long-term relationships with secure partners, which gradually shift dimensional scores toward lower anxiety and lower avoidance; major negative relationship events (betrayal, loss, severe conflict), which can shift scores in the insecure direction; and reflective therapeutic work, particularly approaches that target attachment representations explicitly.
The concept of earned secure attachment — adults who score in the secure range despite adverse early histories — is one of the more important and underreported findings in the literature. Research using the AAI shows that earned-secure adults develop parenting behaviors closely resembling continuously-secure adults, suggesting the pathway is functionally real. The mechanism appears to involve developing a coherent, reflective narrative about adverse early experiences — being able to describe them with clarity and perspective rather than remaining overwhelmed by them or dismissively cutting them off. Some research suggests earned-secure adults retain slightly elevated rates of depression relative to continuously-secure adults, indicating that the pathway is real but does not entirely erase the residue of insecure early experience.
Cross-Cultural Evidence
Attachment theory makes a universality claim: that the attachment behavioral system is a feature of human evolutionary heritage, not a product of Western cultural norms. The cross-cultural evidence broadly supports this claim while also revealing meaningful variation in how insecurity is distributed across cultures.
Van IJzendoorn and Kroonenberg's landmark 1988 meta-analysis synthesized Strange Situation data from 32 samples across eight countries. The headline finding was consistent: secure attachment was the modal pattern in every culture sampled — accounting for roughly 65–70% of infants across samples. The behavioral system appears universal.
The variation was in insecure distributions. German samples showed relatively higher rates of insecure-avoidant attachment compared to American samples; Japanese samples showed relatively higher rates of insecure-resistant attachment. Van IJzendoorn and Kroonenberg noted that within-country variation across studies was actually larger than between-country variation — suggesting that sampling differences, caregiving context, and socioeconomic factors may account for more of the variation than national culture per se.
The Japanese resistant finding has generated particular methodological debate. Japanese caregiving norms involve closer physical proximity between mothers and infants than American norms — less routine separation, more co-sleeping, on-demand care. The Strange Situation's stress procedure (brief lab separations) may therefore be more extreme relative to the infant's ordinary experience in Japan than in the US, producing elevated distress that is scored as resistant but may reflect cultural unfamiliarity with separation rather than genuine insecurity. This interpretation remains contested and illustrates a broader methodological limit: instruments designed in one cultural context may misread normal variation in another.
Cross-cultural adult self-report research generally confirms that the two-dimensional anxiety × avoidance structure holds across diverse samples — the factor structure of instruments like the ECR-R replicates in many populations — while showing meaningful cross-cultural variation in mean levels. Collectivist cultures tend to show somewhat different patterns of avoidance than individualist cultures. The behavioral system appears universal; its average expression varies with cultural context.
Predictive Validity
What does attachment style actually predict, and how strongly?
Relationship outcomes. Insecure attachment — particularly high anxiety — consistently predicts lower relationship satisfaction, both for the individual and (as partner effect) for their partner. A meta-analysis by Zhang et al. (2022) covering 245 samples and N=79,722 participants found that higher attachment anxiety and avoidance were positively associated with negative affect outcomes (depression, anxiety, loneliness) and negatively associated with positive affect outcomes (life satisfaction, self-esteem). The effect size for avoidance and negative mental health was r=.28 — moderate, consistent, and not explained away by Big Five personality traits or other confounders in studies that have tested this.
Mental health. The associations between attachment insecurity and psychopathology are well-documented. High attachment anxiety predicts internalizing disorders (anxiety, depression) more consistently than avoidance does; high avoidance predicts difficulties with emotional intimacy and support-seeking but is less strongly linked to clinical symptom levels, particularly on self-report measures where avoidant individuals tend to minimize distress. Disorganized attachment in infancy is one of the more robust developmental predictors of dissociation, borderline personality features, and anxiety disorders, particularly in samples with histories of early maltreatment.
Parenting and intergenerational transmission. The AAI's ability to predict infant Strange Situation classification — including prenatally — is the most striking predictive finding in the attachment literature. Van IJzendoorn's (1995) meta-analysis of 18 studies (N=854) found approximately 75% concordance between parent AAI classification and infant Strange Situation classification. The mechanism is caregiver sensitivity: secure-autonomous parents show more consistent, sensitive responses to infant attachment signals, which organizes the infant's attachment. Disorganized infant attachment is specifically predicted by unresolved AAI classification in the parent — parents who become confused or distressed when discussing past trauma or loss tend to show frightening or dissociative behavior toward their infants, which creates the irresolvable conflict that produces disorganized infant attachment.
Therapy outcomes. Higher attachment anxiety at the start of therapy predicts stronger alliance formation early in treatment — anxiously attached clients tend to engage intensively with the therapeutic relationship. Higher avoidance predicts more difficulty forming alliance early, but avoidant clients who do form strong alliances show comparable or better outcomes than anxious clients. Attachment insecurity can be reduced through therapy, and movement toward greater attachment security is associated with better treatment outcomes across a range of modalities (Mikulincer & Shaver). The therapeutic relationship itself functions as an attachment relationship — providing the safe haven and secure base that Bowlby described — and this framing has become foundational in contemporary attachment-informed clinical practice.
The Categorical vs. Dimensional Question
A technically important finding that popular attachment content consistently ignores: the four-style categories are useful heuristics, not natural kinds.
Taxometric analyses — statistical procedures that test whether data are better described as categories or dimensions — consistently find that variation in attachment is dimensional, not categorical. The ECR-R authors (Fraley, Waller, & Brennan, 2000) explicitly advise against classifying people based on their continuous ECR-R scores because doing so reduces measurement precision and statistical power. Someone who scores at the exact quadrant boundaries is categorically "secure" but is not the same person as someone who scores near zero on both dimensions — yet the label treats them identically.
This has practical implications for how to read a result. The anxiety and avoidance scores from the assessment are more informative than the style label. A person with moderate anxiety and very low avoidance is different from a person with high anxiety and moderate avoidance, even if both are categorically "anxious-preoccupied." The dimensions capture the difference; the category does not.
The four-style labels are retained because they are useful communication tools — they provide intuitive anchors for the dimensional space, and most therapy and relationship research is organized around them. But they should be understood as approximations rather than natural boundaries.
Limits of the Evidence
Attachment theory is well-supported by decades of research, but several genuine limits are worth naming.
The Western bias in instrument design remains a methodological concern despite cross-cultural replications. The Strange Situation, the AAI, and most self-report instruments were developed in North American and European contexts. Cross-cultural validity has been demonstrated for many populations but cannot be assumed for all.
The retrospective problem in adult self-report research is real. Most evidence linking adult attachment to childhood caregiving is retrospective — adults reporting on their recollections of early relationships. Retrospective reports are affected by current mood, current attachment security, and the tendency to construct coherent narratives about the past. Prospective longitudinal studies exist but are fewer; the correlation between measured infant Strange Situation classification and adult self-reported attachment security is typically modest — one longitudinal study cited by Fraley (2002) found r=.17 over 20 years.
The effect size question: the predictive correlations in attachment research are generally moderate — typically in the .20–.40 range for relationship and mental health outcomes. These are consistent and replicable, but they leave substantial unexplained variance. Attachment style is one determinant of relationship functioning among many, not a master variable.
The AAI-ECR-R gap: the two research traditions do not converge as cleanly as theorists initially hoped. Their cross-classification agreement is imperfect, which means findings from one tradition cannot be automatically generalized to the other. A claim supported by AAI research is not automatically supported by ECR-R research, and vice versa.
For a deeper look at how attachment theory developed — from Bowlby's break with psychoanalysis through Main's Adult Attachment Interview — see the history page. For the four adult attachment styles as they are currently understood, see Secure, Anxious-Preoccupied, Dismissive-Avoidant, and Fearful-Avoidant. To find your own attachment profile, take the assessment.