Bayesian Network Analysis Reveals Relationship Dynamics

Attachment questionnaires ask 36 things and score two: how anxious you are about closeness and how much you avoid it. Lortaraprasert and colleagues (2021) asked a different question: which of those 36 behaviours drive the others? Their answer, a Bayesian network built from a large public dataset of Experiences in Close Relationships (ECR) responses, has a shape worth knowing.

Key takeaways

A Bayesian network is a map of which answers predict which. Two survey items sit at the root, one from the avoidance scale and one from the anxiety scale; everything else flows from them.
The network splits into five clusters: two avoidance patterns and three anxiety patterns. They line up with earlier factor-analysis studies, a good sign the structure is real.
Root behaviours are the leverage points. If a pattern in your texting traces back to one, that is the thing to work on, not the symptoms downstream.
The data is self-report survey answers, not observed behaviour. Treat the map as a guide to where to look in your own messages, not a diagnosis.

What a Bayesian network is

Correlation tells you two things move together. A Bayesian network goes a step further: it fits a directed graph in which each node depends on its parents, so you can read off which behaviours are upstream. An arrow from A to B means B is best explained by A, not the reverse.

The researchers fitted one to the ECR, the standard 36-item questionnaire behind the anxious/avoidant model of attachment, using a large public dataset from the Open Psychometrics project. The output is a directed acyclic graph: arrows, no loops.

"The network can be divided into 5 clusters, 2 avoidance and 3 anxiety clusters. Furthermore, our list of items in the clusters are consistent with the findings of previous factor analysis studies."— Lortaraprasert et al., 2021

What the network shows

Two roots

One avoidance item and one anxiety item have no parents: nothing in the questionnaire explains them, and they explain a great deal downstream. Each dimension of attachment has a single behaviour that acts as its source.

Five clusters

Two avoidance clusters and three anxiety clusters. Within a cluster, items predict each other; across clusters, they barely do. The groupings match earlier factor analyses, and the fitted coefficients correlate with a separate partial-correlation network study.

Five ends

Some items are pure outcomes: predicted by others, predicting nothing. These are the symptoms, the behaviours you notice first and which are least useful to work on directly.

Anxiety has more shapes

Three anxiety clusters to two avoidance ones. In this data, anxious attachment expressed itself in more distinct groupings of behaviour, which fits the experience that anxious patterns are easy to see and hard to summarise.

What this looks like in a text thread

Attachment anxiety and avoidance both show up in messaging, and the network tells you how to read them: look for the upstream behaviour, not the downstream noise. Our guide to attachment styles in digital communication goes through the signals; three examples:

Anxious pattern. Double-texting, reading into reply gaps and asking "are we ok?" are downstream symptoms. The upstream behaviour is the worry itself, which a partner cannot fix with faster replies.
Avoidant pattern. Short replies to long messages, going quiet after intimacy, deflecting into logistics. Also downstream. The upstream behaviour is discomfort with showing feeling, which is why "please text more" rarely works.
Mixed pairs. An anxious cluster in one partner and an avoidant cluster in the other produce the pursue-withdraw loop, visible in a chat as rising message counts on one side and lengthening gaps on the other.

MosaicChats measures the downstream layer: who initiates, how reply times and lengths compare, how tone moves week by week. That is the right place to start, because it is the layer you can see. The network is the reminder that when you find a pattern there, the cause is usually one level up.

The data is self-report questionnaire answers, not observed behaviour, and a network fitted to survey data shows dependence, not proof of cause. The authors' check that the clusters match earlier factor-analytic work is reassuring; it is not a longitudinal study.

See the downstream layer

Upload a chat and MosaicChats shows who initiates, how reply patterns compare and how tone has moved: the behaviours an attachment pattern produces in text.

Analyze your chat

Frequently asked questions

What is a Bayesian network in relationship research?

A directed graph fitted to data in which each behaviour depends on its parents, so you can read off which behaviours are upstream of which. Lortaraprasert et al. (2021) fitted one to attachment questionnaire (ECR) responses and found two root items and five clusters.

What are the five clusters?

Two avoidance clusters and three anxiety clusters. Items within a cluster predict one another, and the groupings match earlier factor-analysis studies of the ECR, which suggests the structure is stable rather than an artefact of one dataset.

Can this predict whether my relationship will last?

No. The study maps how attachment behaviours relate to each other in survey data; it does not follow couples over time. Use it to understand which of your patterns are causes and which are symptoms, not as a forecast.

How does attachment style show up in texting?

Anxious patterns tend to appear as double-texting, reading into reply gaps and reassurance-seeking; avoidant patterns as short replies, going quiet after closeness and deflecting into logistics. Those are downstream behaviours; the network suggests each dimension has a single upstream tendency behind them.

References & Sources

Lortaraprasert, P., Manoret, P., Jantrachotechatchawan, C., & Duangrattanalert, K. (2021). Using Bayesian network analysis to reveal complex natures of relationships. arXiv, 2111.06640. Source

Bayesian NetworksReveal Relationship Dynamics