(PS2-100) Which Possessions Do People Cherish and Why? Using Latent Dirichlet Allocation to Examine Patterns in Object Saving Behaviors and Their Relationship to Object Attachment and Hoarding Symptoms

Hoarding Disorder is characterized in part by object saving behaviors, and the motivations for these behaviors vary widely across individuals (Frost et al., 2015). Research has leveraged etiological theories (e.g., attachment theory) to examine possible motivations for saving behaviors (e.g., object attachment). A bottom-up, data-driven approach may complement top-down, theory-driven approaches and provide novel information about the motivations for saving behaviors that underlie hoarding symptoms. This study used natural language processing to examine data-driven patterns in motivations for object saving behaviors and their relationship to object attachment and hoarding symptoms. 

Participants included 442 adults (M = 38 yrs, 45% male, 73% White, 88% Non-Hispanic) recruited via Prolific. Approximately half of the sample (49%) reported receiving prior mental health treatment. Participants completed the Object Attachment Questionnaire for a personal and non-personal object as well as the Hoarding Rating Scale. Participants were asked to describe a cherished object and an associated memory. Latent Dirichlet Allocation was used with the object/memory descriptions to identify latent topics (i.e., groups of words that tended to appear together in text responses). Model tuning was used to identify the optimal number of topics prior to model identification. Participants were classified into topic groups based on the proportion of words in their responses that belonged to a given topic. Chi-square tests of independence, independent samples t tests, and regression analyses were conducted to compare topic groups on demographics, object attachment, and hoarding symptoms. 

A two-topic solution best fit the data. Topic 1 represented objects with familial and emotional value; some of the most frequent words in this topic included “gift”, “picture”, and “love”. Topic 2 represented objects with social or functional value; some of the most frequent words in this topic included “play”, “use”, and “friend”. Women were more likely to describe objects with familial/emotional value, and men were more likely to describe objects with social/functional value (χ² = 16.56, p < .001). Participants who described objects with familial/emotional value (M = 40) were older on average than those who described objects with social/functional value (M = 37; t = 2.38, p = .018). Topic groups did not differ by race/ethnicity. Participants who described objects with familial/emotional value reported similar attachment to personal objects but greater attachment to non-personal objects than those who described objects with social/functional value (t = 2.02, p = .044). Topic groups did not differ on hoarding symptoms. 

Natural language processing can be used to identify meaningful motivations of object saving behaviors that are related to object attachment, but further refinement is needed to better identify individuals with hoarding symptoms. As Latent Dirichlet Allocation considers frequency of word use but not word meaning, future studies should consider Latent Semantic Analysis as an alternative approach to natural language processing of mental health data to classify grammatically different but semantically similar qualitative data.