Head versus heart: social media reveals differential language of loneliness from depression

Loneliness—an emotionally unpleasant experience when interpersonal relationships are perceived as not meeting expectations—is prevalent in the general population and among people with mental health disorders, impacting psychological functioning and predicting increased morbidity and mortality. Prior research shows loneliness is highly correlated with depression (often r = 0.4–0.6) and is frequently treated as a transdiagnostic factor or subsymptom, potentially overshadowing its unique features. Although both conditions share links to social deficits, maladaptive social cognition, interpersonal distress, and uncontrollable thoughts (worry, rumination), emerging evidence indicates loneliness and depression are related yet distinct constructs with unique symptomatology and cognitive mechanisms. Loneliness may especially involve hypervigilance to social threat and negative expectations about social interactions, whereas depression is not necessarily accompanied by such social expectations. The present study aims to distinguish unique markers of loneliness from depression using psychological assessments (ULS-3 and PHQ-9) and Facebook language, identifying overlapping and unique linguistic correlates and evaluating the predictive utility of language features to inform targeted interventions for loneliness.

- Loneliness is widespread and linked to poorer health outcomes, morbidity, and mortality. - Strong associations between loneliness and depression have been documented (e.g., r ≈ 0.4–0.6), with loneliness considered both a risk factor for and consequence of depression. - Both constructs are associated with social deficits, maladaptive social cognitions, interpersonal distress, and rumination. - Evidence suggests loneliness and depression are separable constructs with unique symptomatology; variance in uncontrolled thoughts/worry may distinguish loneliness from depression. - Loneliness has unique cognitive functions, including hypervigilance to social threats, influencing perceptions of social interactions. - Social media language has proven useful in assessing mental health, including detecting depression and other outcomes, motivating its use to differentiate loneliness from depression.

Design and data sources: Secondary analysis of a larger study that recruited U.S.-based participants via Qualtrics Panels. Participants completed psychological assessments and could share access to their Facebook status updates. Participants: Of 3215 recruited, 3043 passed an attention check; 2986 with sufficient Facebook language (more than 500 words) and complete measures (ULS-3, PHQ-9, age, gender) were included. Demographics (as reported): 61.5% aged approximately 18–34 years; 69.7% female; 63.8% with bachelor’s or higher degree. IRB approvals: original (Purdue University); secondary analysis exempted (University of Pennsylvania). Measures: - Loneliness: UCLA Loneliness Scale—3 items (ULS-3), Likert 1–4, total 3–12. Reliability in sample: α = 0.81. - Depression: Patient Health Questionnaire-9 (PHQ-9), 9 items, 4-point frequency scale. Reliability in sample: α = 0.90. Language features from Facebook: - Corpus: 3,459,854 Facebook posts from included participants. - Closed-vocabulary: LIWC 2015 categories; computed relative frequencies of categories per user. - Open-vocabulary words/phrases: 1–3-grams tokenized with DLATK/tokenization toolkit; tokens used by <1% of users removed, resulting in 4143 unique n-grams. - Topics: Latent Dirichlet Allocation (LDA) topics using an open-source topic set trained on a large corpus of Facebook statuses to obtain user-level topic prevalences. - Contextual embeddings: BERT embeddings (10th layer) averaged to message level then to user level. Analytic approach: - Person-level analyses. For association analyses, each language feature dimension (LIWC category, n-gram, LDA topic) was entered as an independent variable in separate OLS regressions predicting loneliness (ULS-3) and depression (PHQ-9), controlling for age and gender; additional models further controlled for depression (or loneliness) to identify unique associations. Benjamini–Hochberg False Discovery Rate (BH-FDR) corrections applied; Pearson correlations reported with 95% CIs. - Survey correlations among age, loneliness, and depression computed (Pearson) with partial correlations controlling for variables as specified. - Predictive modeling: Compared feature sets (demographics; n-grams; topics; BERT; combinations) for predicting loneliness and depression at the user level, reporting correlation between predicted and observed scores (r) and error metrics as provided.

- Survey correlations: Loneliness and depression were strongly correlated (e.g., r ≈ 0.54–0.55; p < 0.001). Age correlated negatively with both loneliness (r ≈ -0.14) and depression (r ≈ -0.26). Gender showed no significant effects on either outcome (|t| < 1.7, p > 0.1). - Overlap in language markers: A substantial overlap between loneliness- and depression-associated language across methods: - LIWC: 87.1% of loneliness-associated categories also associated with depression (BH-FDR p < 0.001). - Words/phrases: 51.8% overlap (p < 0.05–0.094 threshold indicated). - LDA topics: 75.4% overlap (p < 0.001). - Shared risk markers: more mentions of sickness, pain, negative emotions; cognitive process language (differentiation, tentative, insight, causation); greater first-person singular pronouns; themes of sleep, present orientation, risk, death, intergroup content, and fillers. - Shared protective markers: more references to social relationships and activities (e.g., LIWC we, affiliation; words/phrases like our, birthday, wedding; topics about social gatherings and relationships). - Unique to loneliness after controlling for depression (in addition to age, gender): - Increased cognitive focus: LIWC uncertainty (r = 0.061, 95% CI [0.025, 0.097]), tentative (r = 0.055, [0.019, 0.090]), insight (r = 0.061, [0.026, 0.097]); auxiliary verbs (r = 0.055, [0.020, 0.091]); common adverbs (r = 0.057, [0.022, 0.093]). - Topics reflecting cognition/observation and reading/writing (e.g., poems, writer, publishing) appeared as loneliness risk markers uniquely. - Social relationship language remained protective (e.g., LIWC first-person plural: r = -0.070, [–0.106, –0.035]; LIWC affiliation: r = -0.079, [–0.114, –0.043]). - Several markers of negative emotions, sickness/pain, first-person singular, and sleep lost significance for loneliness once controlling for depression. - Example LIWC associations before controlling for depression: - Risk: Anxiety (r = 0.096, [0.060, 0.131]); Sadness (r = 0.090, [0.054, 0.125]); Anger (r = 0.087, [0.051, 0.122]); Uncertainty (r = 0.107, [0.071, 0.142]); Tentative (r = 0.104, [0.068, 0.139]); Insight (r = 0.095, [0.059, 0.130]); First-person singular (r = 0.101, [0.065, 0.136]). - Protective: First-person plural (r = -0.122, [–0.157, –0.087]); Affiliation (r = -0.113, [–0.148, –0.078]); Friends (r = -0.060, [–0.095, –0.024]). - Predictive performance: Facebook language features predicted both outcomes and outperformed demographics alone; prediction was weaker for loneliness than depression. - Loneliness: Age+Gender r = 0.133; BERT r = 0.201. - Depression: Age+Gender r = 0.253; BERT r = 0.312.

Findings show substantial overlap in language markers between loneliness and depression, indicating shared risk (sickness, pain, negative affect, cognitive process language, first-person singular) and protective (social relationships/activities, first-person plural, affiliation) linguistic features. Crucially, after controlling for depression, loneliness exhibits a stronger cognitive signature: greater emphasis on uncertainty, tentative and insight language, observation, reasoning, and self-directed cognitive activities (e.g., reading, writing). In contrast, depression is more characterized by negative emotions, pain perception, and affect-focused rumination. This “head versus heart” distinction suggests that loneliness more strongly involves maladaptive social cognition and overattention to environmental interpretation, whereas depression more heavily reflects affective distress. These insights suggest interventions for loneliness should prioritize modifying maladaptive social cognitions (e.g., cognitive reframing of social environments), enhancing and repairing social relationships, and concurrently addressing comorbid depressive symptoms when present. The predictive analyses demonstrate the utility of social media language in assessing loneliness and depression at scale, though loneliness is somewhat harder to predict than depression, reflecting its nuanced and socially contextual nature.

This study leverages large-scale Facebook language linked with validated assessments (ULS-3, PHQ-9) to disentangle shared and unique linguistic markers of loneliness and depression. While many markers overlap, loneliness shows a distinct cognitive focus—uncertainty, tentative/insightful language, observation, and reading/writing—beyond depression’s core affective signature. Social relationship language remains protective for loneliness even after controlling for depression. Language features can predict both loneliness and depression, though with lower accuracy for loneliness. These findings inform interventions that target maladaptive social cognitions, strengthen social ties, and address co-occurring depressive affect. Future research should employ longitudinal designs, broader age ranges (including adolescents and older adults), more comprehensive loneliness measures (e.g., UCLA-20), and time-aligned language to capture the episodic dynamics of loneliness and depression.

- Correlational design precludes causal inference between language features and loneliness/depression. - U.S. adult sample limits generalizability to adolescents and older adults, who may experience loneliness differently. - Loneliness assessed with a brief 3-item scale, limiting precision and predictive power relative to longer instruments. - Single-time assessment of loneliness and depression contrasted with language spanning long periods; temporal misalignment may obscure episodic changes. - Some reported methodological details (e.g., topic model source/scale) and table formatting suggest potential transcription irregularities in the text; however, core findings are robust across methods and corrections.