Loneliness Biases Social Engagement Decisions: Evidence From Value‐Based Choice Models

ABSTRACT

Loneliness is a well‐established risk factor for adverse mental health outcomes and persistent social withdrawal. We tested whether loneliness is associated with systematic alterations in value‐based social decision‐making and examined mechanisms that may sustain withdrawal. Across two preregistered online studies (total N = 6517), participants were randomly assigned to positive, negative, or neutral social expectation priming and completed a vignette‐based task manipulating reward, effort, and threat. Loneliness and psychopathological symptoms were assessed by self‐reports. Across both studies, higher loneliness was consistently associated with a lower likelihood of engaging in social interactions, independent of vignette characteristics and expectation priming. Higher loneliness was consistently associated with lower anticipated connectedness and greater perceived rejection risk during social decision‐making. Further exploratory analyses indicated that loneliness was associated with reduced sensitivity to both social reward and social threat, accompanied by diminished differentiation between high‐ and low‐reward contexts, consistent with a compressed range of subjective social value across decision scenarios. These findings suggest that loneliness is linked to altered social valuation, characterized by blunted differentiation of social rewards and reduced motivational salience of potentially beneficial interactions, contributing to persistent withdrawal across diverse everyday social contexts.

In two preregistered studies, participants first completed an expectation priming task (positive, neutral, or negative) and subsequently made decisions in a vignette‐based social decision task in which potential reward, required effort, and social threat were systematically manipulated. Exploratory analyses indicated that higher loneliness was associated with attenuated sensitivity to both rewarding and threatening social cues. Furthermore, loneliness was linked to diminished differentiation between high‐ and low‐reward contexts, reflecting a narrowed range of subjective social value across decision scenarios.

1. Introduction

Loneliness is commonly defined as a subjective experience arising from a discrepancy between desired and perceived social relationships [1]. It reflects the perceived quality and value of social connections rather than their quantity, and is, therefore, conceptually distinct from social isolation [1, 2]. Persistent loneliness is associated with adverse mental and physical health outcomes, underscoring its relevance as a public health concern [3, 4, 5, 6]. These consequences highlight the need to understand the mechanisms through which loneliness shapes behavior and to inform interventions.

Contemporary theories increasingly conceptualize loneliness not merely as a lack of social contact, but as a maladaptive cognitive‐motivational state that biases social decision‐making by altering how social options are evaluated and selected [2, 7]. Lonely individuals tend to anticipate rejection, discount social rewards, and overestimate engagement costs, shifting choices toward withdrawal [8]. Models emphasizing pessimistic social expectations propose that loneliness fosters negative anticipations regarding social encounters, which in turn shape subsequent social decisions [9]. Despite substantial correlational evidence documenting cognitive and affective biases associated with loneliness, direct causal evidence linking loneliness to social decision‐making remains limited [9]. Much experimental research on loneliness has relied on self‐report measures, abstract judgments about hypothetical social situations, or static vignette paradigms, often indexed using broad measures of social approach and avoidance. However, most prior paradigms assess global evaluations rather than concrete social choices.

Critically, such paradigms rarely require participants to make explicit choices between social options that differ in anticipated reward, required effort, and social risk. In everyday social life, engagement involves repeated value‐based decisions in which potential benefits (e.g., pleasure or connectedness), costs (e.g., effort or time investment), and risks (e.g., rejection or exclusion) must be weighed against one another [10, 11]. Decision‐making research shows that these components can exert dissociable and sometimes opposing effects on choice behavior, and that individual differences can selectively alter sensitivity to specific decision components rather than uniformly shifting approach or avoidance tendencies [12, 13]. However, loneliness research has rarely used paradigms that independently manipulate these components, limiting mechanistic insight into biased social choice.

Consequently, it remains unclear which elements of social valuation are most affected by loneliness. Loneliness may heighten sensitivity to social threat [14], blunt sensitivity to potential social reward [15], inflate perceived effort associated with engagement [16], or alter how these factors are integrated during choice. Distinguishing between these possibilities is critical for understanding the mechanisms through which loneliness translates into reduced social engagement and persistent withdrawal. A complementary limitation concerns the causal role of social expectations. Research on social cognition suggests that recalling prior social experiences can shape expectations about future interactions: reflecting on positive encounters increases anticipated reward and reduces perceived risk, whereas reflecting on negative encounters has the opposite effect [17]. To our knowledge, no prior studies have experimentally manipulated social expectations to test whether altering social forecasts changes engagement decisions, particularly in interaction with individual differences in loneliness. More broadly, these gaps raise the possibility that loneliness may not simply bias specific components of social evaluation, such as threat or reward, nor depend primarily on transient social expectations. Instead, loneliness may operate by compressing the overall landscape of subjective social value, reducing the relative advantage of highly rewarding social opportunities over less rewarding alternatives. Such compression would reduce motivation to engage even when incentives are favorable, reframing loneliness as a disruption of social valuation rather than uniform avoidance.

To address existing research gaps, we conducted a large‐scale, preregistered online experiment in which participants repeatedly decided whether to engage in brief, realistic social interactions that systematically varied in potential reward, required effort, and social threat. Prior to the task, participants were randomly assigned to positive, negative, or neutral social expectancy priming conditions designed to bias expectations regarding future social interactions. This allowed us to test whether altering social expectations influences engagement decisions and whether loneliness‐related valuation processes are context‐sensitive. Given that symptoms of depression [18], anxiety [6], and paranoia [19] are closely associated with loneliness and may independently bias social decision‐making, psychopathological measures were included as covariates in sensitivity analyses to ensure that the observed associations between loneliness and social engagement were not attributable to comorbid psychopathology. Based on prior theory, we preregistered confirmatory hypotheses (Table S1) concerning the main effect of loneliness on engagement, its modulation by social threat and expectation priming, the independent effects of reward, effort, and threat on choice, and the associations of loneliness with proximal social appraisals, including anticipated connectedness, perceived rejection risk, anticipated pleasure, and perceived effort. In addition, we conducted exploratory analyses to characterize how loneliness relates to the valuation of social opportunities, including individual differences in sensitivity to rewards, costs, and threats and in the differentiation of social value across options.

2. Materials and Methods

2.1. Participants and Procedures

A detailed preregistered protocol is available at the Open Science Framework (https://osf.io/jm672/overview). The study was conducted online using two independent Polish research panels. The study was intentionally conducted in two independent panels (“Pollster”—study 1 and “Ariadna”—study 2) using identical procedures to enable internal replication and evaluation of effect robustness across samples. Both panels maintain large pools of preregistered and verified adult volunteers. The two studies were conducted approximately 1 month apart, reducing the likelihood of duplicate participation. Additionally, at the end of each study, participants were asked two self‐report questions—whether they were registered in the opposite panel and whether they had previously taken part in a similar study. Those who responded affirmatively to both questions were removed during data curation.

Quota‐based sampling was applied to approximate the demographic structure of the Polish general population with respect to age, sex, and region of residence. Specifically, recruitment targets were defined for predefined demographic subgroups, and data collection within each subgroup was discontinued once the assigned quota was reached, ensuring that the final sample mirrored population proportions derived from national census data. Eligibility criteria included age (≥18 years), residency in Poland, and Polish language fluency. Panel providers distributed invitations containing a study description, estimated duration (about 25–30 min), compensation information, and a unique survey link. Exclusion criteria included incomplete interviews, eligibility screen‐outs (i.e., participants not meeting age, residency, or language criteria), quota fulfillment (i.e., discontinuation of enrollment once predefined demographic targets were reached), failures of data quality procedures, and duplicate participation risk. Data quality procedures were implemented during data collection and preprocessing. Participants were excluded if they completed the survey unusually quickly (i.e., in less than 30% of the median completion time), failed embedded attention‐check items assessing task engagement, or provided implausible, nonsensical, or randomly generated text responses. The panels differed in the type of attention check items used and their placement within the survey; however, they were comparable in number. At the end of each study, participants were asked two self‐report questions—whether they were registered in the opposite panel and whether they had previously taken part in a similar study. Those who responded affirmatively to both questions were removed during data curation.

A mixed factorial experimental design combined between‐subject and within‐subject manipulations. Participants were randomly assigned to one of three priming conditions—positive, neutral, or negative social expectancy—implemented via a brief writing task in which participants recalled and described a positive, negative, or neutral social situation (minimum length: five sentences or 200 characters, depending on panel constraints). Minimum length requirements were enforced automatically by each platform according to its technical capabilities: one panel supported sentence‐level validation (minimum: five sentences, study 1), while the other supported only character‐level validation (minimum: 200 characters, study 2). Despite this minor procedural difference, the core task was identical across both panels. To verify the effectiveness of the priming manipulation, participants completed six post‐writing ratings assessing the vividness and perceived accuracy of the recalled situation, positivity of the recalled event, perceived acceptance and support from others, current positive affect, and expectations regarding future social situations. All items were rated on 7‐point scales.

Participants were subsequently presented with 12 standardized social decision vignettes varying in effort (low, medium, high), reward (low, high), and social threat (absent, present). These factors were fully crossed (3 × 2 × 2), yielding 12 unique vignette types. After each vignette, participants indicated whether they would engage in the social interaction (join vs. decline; primary outcome) and rated secondary outcomes—including confidence, anticipated pleasure, connectedness, perceived effort, and perceived risk—on 1–7 scales. Vignette order was randomized independently for each participant. An example vignette is as follows: “An acquaintance invites you to meet at a café 20 min from your home. These meetings are usually brief, and the conversation tends to stay superficial.” (medium effort, low‐reward, and no threat). The full set of vignettes (English and Polish versions) is available in the Supporting Information. Participants in study 1 received financial compensation of approximately 10 EUR for participation. Participants in study 2 received standard panel‐based compensation in the form of redeemable points consistent with the provider's reward policy. The study received approval of the Bioethics Committee at Wroclaw Medical University, Wroclaw, Poland. All participants signed a digitalized version of informed consent.

2.2. Measures

After completing the priming and vignette‐based tasks, participants completed a set of standardized questionnaires (Table S2). These included the De Jong Gierveld Loneliness Scale–11 (DJGLS‐11) [20, 21], the Lubben Social Network Scale–6 (LSNS‐6) [22], the Patient Health Questionnaire–9 (PHQ‐9) [23], the Generalized Anxiety Disorder–7 (GAD‐7) [24], the Social Interaction Anxiety Scale (SIAS‐20) [25], and the Revised Green et al. Paranoid Thoughts Scale (R‐GPTS) [26, 27]. Internal consistency for all questionnaires was high across both studies (Cronbach's α values ranged from 0.83 to 0.97), indicating good‐to‐excellent reliability.

2.3. Power Simulations

The target sample size was 3000 participants per panel (two panels; total N = 6000) to ensure precise estimation of effects, stable estimation in repeated‐measures models, and internal replication across independent samples. A priori Monte Carlo simulations using logistic generalized estimating equations (12 trials per participant; realistic fixed effects and random‐intercept variability) indicated that this sample size would provide adequate sensitivity to detect relatively small preregistered interaction effects, including the loneliness × threat interaction corresponding to an OR of 0.90. Sensitivity analyses further indicated substantially lower sensitivity for smaller effects (e.g., OR = 0.95), supporting the rationale for the planned sample size. These estimates were robust across baseline engagement probabilities ranging from 0.50 to 0.55.

2.4. Data Analysis

All preprocessing procedures and analyses were implemented in R (version 4.4.0) using the packages lme4, lmerTest, dplyr, tidyr, and ggplot2 (analytical codes are available in the Supporting Information, and raw data supporting the findings can be found at https://osf.io/jm672/overview). The analyses were conducted separately for study 1 and study 2, following identical analytic pipelines to allow direct assessment of the consistency and reproducibility of findings rather than relying exclusively on pooled estimates across panels. Dummy coded variables included decision to engage (0 = no, yes = 1), threat (0 = absent, 1 = present), and reward (0 = low, 1 = high). Effort was treated as a categorical ordinal variable with three levels (low vs. medium vs. high). Priming condition was a between‐subjects categorical variable with three levels: neutral (reference condition), positive, and negative. All continuous variables (the scores derived from questionnaires measuring symptoms and social network as well as age) were z‐standardized before data analyses. Participant ID and vignette ID were treated as grouping variables for random effects.

To compare general sample characteristics between study 1 and study 2, continuous variables (e.g., age, loneliness scores, and task ratings) were compared using independent‐samples t‐tests, and categorical variables (e.g., gender, education) were compared using χ² tests of independence. Effect sizes are reported as Cohen's d for continuous variables and Cramér's V for categorical variables. To verify successful randomization and to examine baseline differences across priming conditions (neutral, positive, and negative), we compared demographic variables, loneliness scores, and baseline task measures using one‐way ANOVAs for continuous outcomes and χ² tests for categorical outcomes. When omnibus tests were significant, pairwise comparisons were conducted with corrections for multiple comparisons (the Games−Howell test). These analyses were conducted separately for each study. The level of significance was set at p < 0.05.

Primary hypotheses were tested using generalized linear mixed‐effects models (GLMMs) with a logit link function. The dependent variable was decision to engage. Fixed effects included the level of loneliness, threat, reward, effort, and priming condition. Random intercepts were included for participants and vignette types. The primary confirmatory model tested main effects and preregistered interactions using a stepwise model‐comparison strategy: (1) main‐effects model (loneliness, threat, reward, effort, and priming); (2) two‐way interaction model, adding loneliness × threat and loneliness × priming interactions; and (3) three‐way interaction model, adding loneliness × threat × priming. Nested models were compared using likelihood‐ratio tests (χ²). Multiple testing was not accounted for because the hypotheses were preregistered, theory‐driven, and tested within a single, hierarchical model. Covariates were included in sensitivity analyses (symptom scores and social network size were included in independent GLMMs). Additional sensitivity analyses were performed for variables showing significant differences between groups assigned to distinct priming conditions. The level of significance was set at p < 0.05.

Secondary confirmatory hypotheses were examined using linear mixed‐effects models. Separate models were estimated for each proximal appraisal outcome (anticipated connectedness, perceived rejection risk, anticipated pleasure, and perceived effort), with participant and vignette specified as random intercepts to account for repeated observations and stimulus clustering. Fixed effects included loneliness, vignette characteristics (reward, threat, and effort), and priming condition. Multiple testing across the family of secondary hypotheses was controlled using the Benjamini–Hochberg false discovery rate (FDR) procedure.

To descriptively characterize latent valuation processes underlying social decisions, further exploratory analyses were conducted. These analyses were not preregistered and should be interpreted as hypothesis‐generating. Participants’ binary engagement decisions were modeled within a logistic choice framework, in which the probability of engagement was determined by a latent subjective value signal derived from vignette attributes. Subjective value was specified as a linear combination of reward, threat, and effort attributes, which were contrast‐coded and mean‐centered (reward: high = +0.5, low = −0.5; threat: present = +0.5, absent = −0.5; effort: low = −0.5, medium = 0, high = +0.5). This additive structure was specified a priori based on the experimental manipulation of reward, threat, and effort dimensions and was intended to provide an interpretable summary of individual differences in social valuation sensitivity rather than to identify an optimal predictive model of choice behavior. Individual differences in sensitivity to each attribute were estimated using GLMMs with participant‐level random intercepts and random slopes for reward, threat, and effort. Participant‐specific sensitivity parameters (empirical Bayes estimates) were extracted and examined in secondary linear models to assess associations with z‐standardized loneliness and priming condition. To facilitate comparison across parameters, sensitivity estimates were additionally standardized. To complement parameter‐based analyses, value‐based sensitivity indices were derived, including (i) reward value differentiation, defined as the difference in model‐implied subjective value between high‐ and low‐reward conditions, and (ii) the overall subjective value range across all vignette combinations. These indices were regressed on loneliness to assess whether loneliness was associated with compression of the social value landscape. For visualization only, predicted subjective values were plotted across reward, threat, and effort combinations, stratified by loneliness quartiles. Statistical significance was evaluated at p < 0.05, with Benjamini–Hochberg FDR correction applied across the family of exploratory analyses. Additional sensitivity analyses further adjusted these models for depressive symptoms, generalized anxiety symptoms, social anxiety symptoms, and paranoid ideation. Multicollinearity in all regression models was assessed using the variance inflation factors (VIFs), with values below 5 considered indicative of acceptable levels of multicollinearity [28].

3. Results

3.1. General Characteristics of Participants

A total of 3139 and 3378 participants were finally included in study 1 and study 2, respectively (Table 1 and Figure S1). There were significant differences in age, education, place of residence, employment status, relationship status, social network size, and symptom scores between participants from both studies. Effect sizes were small (Cramér's V = 0.013–0.090; Cohen's d = 0.039–0.149).

TABLE 1.

The comparison of participants between studies.

	Study 1 (n = 3139)	Study 2 (n = 3378)	ES a	p
Age	45.4 ± 15.8	47.8 ± 16.1	0.149	< 0.001
Gender
Men	1462 (46.6)	1618 (47.9)	0.013	0.551
Women	1673 (53.3)	1756 (52.0)
Other	4 (0.1)	4 (0.1)
Education
Primary	96 (3.0)	56 (1.7)	0.090	< 0.001
Vocational	338 (10.8)	233 (6.9)
Secondary	1267 (40.4)	1342 (39.7)
Higher	1438 (45.8)	1747 (51.7)
Place of residence
Rural	1174 (37.4)	1287 (38.1)	0.069	< 0.001
Urban, < 50,000 inhabitants	609 (19.4)	820 (24.3)
Urban, 50,000–150,000 inhabitants	452 (14.4)	442 (13.1)
Urban, 150,000–500,000 inhabitants	430 (13.7)	402 (11.9)
Urban, > 500,000 inhabitants	474 (15.1)	427 (12.6)
Employment status
Unemployed	282 (9.0)	222 (6.6)	0.080	< 0.001
Student	156 (5.0)	149 (4.4)
Part‐time employment	321 (10.2)	246 (7.3)
Full‐time employment	1747 (55.7)	1949 (57.7)
Retired or disability pension	633 (20.2)	812 (24.0)
Relationship status
Single	687 (21.9)	634 (18.8)	0.055	< 0.001
Divorced	238 (7.6)	208 (6.2)
Widowed	158 (5.0)	170 (5.0)
Informal relationship	491 (15.6)	524 (15.5)
Formal relationship	1565 (49.9)	1842 (54.5)
Monthly income
<750 USD	686 (21.9)	791 (23.4)	0.023	0.496
750–1500 USD	1581 (50.4)	1632 (48.3)
1500–2500 USD	670 (21.3)	728 (21.6)
2500–3750 USD	144 (4.6)	161 (4.8)
>3750 USD	58 (1.8)	66 (2.0)
GAD‐7	6.0 ± 5.5	5.4 ± 5.4	0.114	< 0.001
SIAS	27.2 ± 15.1	26.6 ± 14.6	0.043	0.085
PHQ‐9	7.1 ± 6.0	6.5 ± 6.0	0.098	< 0.001
R‐GPTS	16.4 ± 16.8	17.7 ± 16.7	0.073	0.003
DJGLS‐11	5.1 ± 3.8	5.0 ± 3.9	0.039	0.115
LSNS‐6	15.0 ± 6.3	15.9 ± 6.1	0.135	< 0.001

Note: Data reported as mean ± SD or n (%).

Abbreviations: DJGLS‐11, De Jong Gierveld Loneliness Scale‐11; GAD‐7, Generalized Anxiety Disorder‐7; LSNS‐6, Lubben Social Network Scale‐6; PHQ‐9, Patient Health Questionnaire‐9; R‐GPTS, Revised Green et al. Paranoid Thoughts Scale; SIAS, Social Interaction Anxiety Scale.

^a Effect size estimate estimated as Cohen's d (continuous variables) or Cramer's V (categorical variables).

3.2. Participants Across Priming Conditions

Across both studies, participants assigned to the positive, neutral, and negative priming conditions were well balanced on demographic, social, and clinical characteristics (Tables S3 and S4). In study 1, no significant differences were observed across priming conditions for age, gender, education, income, relationship status, loneliness, social network size, or symptoms of depression, anxiety, or paranoia (p > 0.05). In study 2, priming groups were likewise comparable on all baseline measures, with the exception of a small difference in age (p = 0.021).

Manipulation checks demonstrated significant effects of priming on proximal social–emotional appraisals in both samples (Tables S5 and S6). For all appraisal measures, omnibus tests revealed robust differences between priming conditions, with positive social expectancy consistently associated with more positive appraisals and negative social expectancy with less positive appraisals. The neutral condition generally fell between these extremes. Ratings of vividness and perceived accuracy also differed by priming condition, but remained sufficiently high across conditions to indicate that the imagined scenarios were experienced as credible.

3.3. Social Engagement Decisions

Across both studies, GLMMs revealed robust and replicable effects of loneliness and vignette characteristics on the likelihood of engaging in a social interaction (Table 2). In the main‐effects model (model 1), higher loneliness was associated with a lower probability of engagement in both study 1 (OR = 0.800, 95% CI: 0.752–0.851, p < 0.001) and study 2 (OR = 0.758, 95% CI: 0.712–0.807, p < 0.001). Independent of loneliness, the presence of social threat significantly reduced engagement (study 1: OR = 0.365, 95% CI: 0.198–0.670, p = 0.001; study 2: OR = 0.382, 95% CI: 0.208–0.700, p = 0.002), whereas high reward strongly increased engagement, with low‐reward scenarios being associated with substantially lower odds of engagement in both studies (OR = 0.047 in both studies, p < 0.001). Effort exerted a nonlinear influence on decisions, as indicated by a significant quadratic effect in both studies (study 1: OR = 0.475, 95% CI: 0.280–0.805, p = 0.006; study 2: OR = 0.532, 95% CI: 0.315–0.899, p = 0.019), reflecting reduced engagement at higher levels of effort. Priming condition showed no consistent main effect on engagement, although a small effect of positive priming was observed in study 2 only (OR = 1.196, 95% CI: 1.028–1.393, p = 0.021).

TABLE 2.

Mixed‐effects logistic regression predicting social engagement.

Model	Predictor	Study 1				Study 2
		OR	95% CI	VIF	p	OR	95% CI	VIF	p
1a	Loneliness	0.800	0.752–0.851	1.00	< 0.001	0.758	0.712–0.807	1.00	< 0.001
	Threat (present)	0.365	0.198–0.670	1.00	0.001	0.382	0.208–0.700	1.00	0.002
	Priming: positive	1.126	0.969–1.309	1.00	0.120	1.196	1.028–1.393	1.00	0.021
	Priming: negative	1.038	0.893–1.206	—	0.630	0.975	0.838–1.135	—	0.746
	Reward (low)	0.047	0.025–0.086	1.00	< 0.001	0.047	0.026–0.086	1.00	< 0.001
	Effort (linear)	0.667	0.394–1.131	1.00	0.132	0.672	0.398–1.136	1.00	0.138
	Effort (quadratic)	0.475	0.280–0.805	—	0.006	0.532	0.315–0.899	—	0.019
2b	Loneliness	0.891	0.796–0.997	1.48	0.045	0.779	0.699–0.868	1.73	< 0.001
	Threat (present)	0.351	0.191–0.647	1.01	< 0.001	0.380	0.206–0.699	1.01	0.002
	Priming: positive	1.067	0.902–1.263	1.12	0.451	1.173	0.991–1.388	1.11	0.064
	Priming: negative	1.028	0.868–1.218	—	0.746	0.991	0.837–1.172	—	0.913
	Reward (low)	0.047	0.025–0.086	1.00	< 0.001	0.047	0.026–0.086	1.00	< 0.001
	Effort (linear)	0.667	0.394–1.129	1.00	0.132	0.672	0.398–1.136	1.00	0.137
	Effort (quadratic)	0.474	0.280–0.803	—	0.006	0.532	0.315–0.899	—	0.019
	Loneliness × threat	0.950	0.898–1.005	1.12	0.072	0.991	0.939–1.046	1.11	0.738
	Loneliness × positive priming	0.938	0.806–1.092	1.40	0.410	0.917	0.788–1.068	1.29	0.264
	Loneliness × negative priming	0.847	0.729–0.984	—	0.030	1.017	0.874–1.184	—	0.827
	Threat × positive priming	1.104	0.963–1.265	1.27	0.156	1.040	0.912–1.185	1.11	0.560
	Threat × negative priming	1.022	0.892–1.172	—	0.752	0.978	0.858–1.113	—	0.733
3c	Loneliness	0.893	0.791–1.008	1.57	0.068	0.789	0.703–0.885	1.84	< 0.001
	Threat (present)	0.351	0.190–0.647	1.01	< 0.001	0.380	0.207–0.700	1.01	0.002
	Priming: positive	1.067	0.902–1.263	1.12	0.451	1.173	0.991–1.389	1.11	0.063
	Priming: negative	1.028	0.868–1.217	—	0.751	0.990	0.837–1.172	—	0.910
	Reward (low)	0.047	0.025–0.086	1.00	< 0.001	0.047	0.026–0.086	1.00	< 0.001
	Effort (linear)	0.667	0.394–1.130	1.00	0.131	0.672	0.398–1.136	1.00	0.138
	Effort (quadratic)	0.474	0.280–0.804	—	0.005	0.532	0.315–0.899	—	0.018
	Loneliness × threat	0.947	0.858–1.044	1.56	0.273	0.969	0.887–1.060	1.83	0.493
	Loneliness × positive priming	0.922	0.777–1.094	1.56	0.351	0.912	0.771–1.080	1.43	0.285
	Loneliness × negative priming	0.856	0.723–1.014	—	0.071	0.984	0.832–1.164	—	0.852
	Threat × positive priming	1.104	0.963–1.265	1.13	0.156	1.039	0.911–1.184	1.11	0.567
	Threat × negative priming	1.023	0.892–1.173	—	0.745	0.979	0.859–1.115	—	0.750
	Loneliness × threat × positive priming	1.031	0.898–1.184	1.56	0.663	1.009	0.885–1.151	1.42	0.890
	Loneliness × threat × negative priming	0.981	0.856–1.125	—	0.788	1.061	0.932–1.208	—	0.372

Note: Adding two‐way and three‐way interactions was not associated with significant improvement of model fit (study 1: p = 0.078 for the comparison of model 1 vs. model 2, p = 0.781 for the comparison of model 2 vs. model 3; study 2: p = 0.726 for the comparison of model 1 vs. model 2, p = 0.653 for the comparison of model 2 vs. model 3). VIF values for categorical predictors represent adjusted generalized variance inflation factors [GVIF^(1/(2×Df))] computed for the overall model term. Therefore, a single value is reported for each categorical predictor.

^a Main effects.

^b Main effects and two‐way interactions.

^c Main effects, two‐way and three‐way interactions.

Adding preregistered two‐way interactions between loneliness, threat, and priming (model 2) did not significantly improve model fit relative to the main‐effects model in either study. No preregistered interaction effects consistently replicated across studies. Inclusion of three‐way interactions (model 3) did not improve model fit, and none of the higher‐order interaction terms reached significance in either study. Accordingly, inference focused on the parsimonious main‐effects model.

The robustness of the main findings was examined while accounting for age in study 2 (Table S7), social network size (in both studies; Table S8), and psychopathological symptoms (in both studies; Table S9). The findings in study 2 were robust to adjustment for age. Adjustment for social network size attenuated the effects of loneliness in both studies (model 1). Across both studies, the findings were robust to adjustment for psychopathological symptoms.

3.4. Proximal Appraisal Analyses

Mixed‐effects models showed that higher loneliness was consistently associated with lower anticipated connectedness and anticipated pleasure, as well as greater perceived rejection risk and perceived effort across both studies (Table S10). Positive priming was associated with more favorable appraisals in some models, particularly in study 2, whereas negative priming showed less consistent effects across samples.

3.5. Loneliness and Value‐Based Sensitivity Parameters

Across both studies, loneliness was robustly associated with altered sensitivity to social reward and threat, but not effort (Figure 1 and Table S11). In both studies, higher loneliness predicted lower reward sensitivity (study 1: β = −0.097, P _FDR < 0.001; study 2: β = −0.156, P _FDR < 0.001) and lower threat sensitivity (study 1: β = −0.063, P _FDR < 0.001; β = −0.102, P _FDR < 0.001). The association with effort sensitivity was observed only in study 2 (β = 0.098, P _FDR < 0.001). In both studies, the strongest effects of loneliness were found for reward sensitivity. Priming showed no consistent association with value‐based sensitivity parameters (Table S11).

FIGURE 1.

Standardized effects of loneliness on model‐derived sensitivity parameters derived from mixed‐effects decision models in study 1 (A) and study 2 (B). Bars represent standardized regression coefficients (β) linking loneliness to individual‐level sensitivity to reward, threat, and effort, estimated from participant‐specific random slopes. Across both studies, higher loneliness was associated with reduced sensitivity to reward and threat, whereas effects on effort sensitivity were not consistent across studies. Asterisks indicate statistically significant associations (p < 0.05).

In both studies, higher loneliness was associated with reduced reward value differentiation (study 1: β = −0.098, SE = 0.018, P _FDR < 0.001; study 2: β = −0.133, SE = 0.015, P _FDR < 0.001; Figure 2 and Table S12), indicating that lonelier individuals assigned more similar subjective values to high‐ and low‐reward options. This effect was significant after adjustment for depressive symptoms, social anxiety, generalized anxiety, and paranoid ideation (study 1: β = −0.062, SE = 0.006, P _FDR < 0.001; study 2: β = –0.111, SE = 0.006, P _FDR < 0.001). Consistent with this pattern, loneliness also predicted a narrower overall subjective value range across conditions (study 1: β = −0.055, SE = 0.018, P _FDR = 0.005; study 2: β = −0.156, SE = 0.020, P _FDR < 0.001), even after accounting for depressive symptoms, social anxiety, generalized anxiety, and paranoid ideation (study 1: β = −0.028, SE = 0.006, P _FDR < 0.001; study 2: β = –0.080, SE = 0.006, P _FDR < 0.001) suggesting a general compression of value representations.

FIGURE 2.

Loneliness‐related differences in reward value differentiation and subjective value curves. Associations between loneliness (z‐score) and reward value differentiation, defined as the difference in subjective value between high‐ and low‐reward options (ΔV = V_high − V_low), are shown in panels A (study 1) and C (study 2). Each point represents an individual participant; the solid line indicates the fitted linear regression. Higher loneliness was associated with reduced reward value differentiation in both studies. Subjective value curves illustrating mean subjective value (V, log‐odds scale) across effort levels (low, medium, high) are shown separately by reward level (high vs. low; column facets) and threat presence (absent vs. present; row facets) in panels B (study 1) and D (study 2). Lines indicate loneliness quartiles (1 = lowest, 4 = highest loneliness). Across both studies, higher loneliness was associated with an overall downward shift in subjective value across conditions, consistent with attenuated reward valuation rather than altered effort‐related curvature.

4. Discussion

Across two large preregistered studies, we showed that loneliness is robustly associated with reduced willingness to engage in social interactions across a wide range of decision contexts. This association was not confined to situations involving elevated social threat, nor did it depend strongly on experimentally induced shifts in social expectations. However, the association between loneliness and engagement was no longer significant after accounting for objective social network size, indicating substantial overlap between perceived social disconnection and the availability of social ties at the level of behavioral choice. Instead, loneliness was characterized by altered evaluation of social opportunities, reflected in less favorable appraisals, reduced reward sensitivity, and diminished value differentiation. Higher loneliness was consistently associated with lower anticipated connectedness and greater perceived rejection risk during social decision‐making. These findings indicate that loneliness is accompanied by less favorable appraisals of potential social interactions, although the present design does not permit conclusions regarding causal pathways linking these appraisals with engagement decisions. These findings align with and extend prior work showing that loneliness is associated with altered sociocognitive processing, including atypical and unstable expectations about others’ emotional behavior, heightened rejection sensitivity, increased attention to social threat, and negative or hostile intent attributions [9, 29, 30, 31, 32]. While prevailing accounts emphasize heightened vigilance for social threat and pessimistic expectations of rejection [33], our results suggest that loneliness is not marked by uniform avoidance of threatening contexts. Rather, it involves a broader disruption in social valuation, characterized by blunted responsiveness to positive social cues and a reduced capacity to discriminate social value across contexts.

Our exploratory analyses provide convergent evidence that loneliness reshapes social valuation by compressing the subjective value landscape of social opportunities. Specifically, the observed compression of subjective value—wherein highly rewarding social options lose their relative advantage over low‑reward alternatives—parallels recent evidence that neural sensitivity to social reward is altered in individuals experiencing loneliness. For example, in youth, variability in neural reward responsiveness was found to predict how social context relates to loneliness and satisfaction, suggesting that individual differences in reward sensitivity modulate social engagement outcomes [34]. This diminished differentiation in value may undermine the motivational salience of social opportunities even when objective incentives are high. Conceptually, this compressed value landscape aligns with predictive processing frameworks positing that loneliness emerges from mismatches between expected and experienced social outcomes, leading to rigid internal models that undervalue affiliative potential [35]. By linking individual differences in sensitivity to social reward and threat with engagement decisions, our results provide quantitative support for theoretical accounts, suggesting that loneliness involves altered social valuation processes that may contribute to persistent social withdrawal. Importantly, the associations between loneliness and altered social valuation remained significant after adjustment for depressive symptoms, generalized anxiety symptoms, social anxiety symptoms, and paranoid ideation, suggesting that these effects were not fully explained by broader psychopathological symptom dimensions.

Although classical formulations of loneliness often emphasize heightened vigilance for social threat as a key driver of withdrawal and stress responses, empirical findings have been mixed, with some work showing elevated threat sensitivity and others suggesting more nuanced patterns [36]. Crucially, our results extend this literature by showing that loneliness is associated with reduced sensitivity to both positive (rewarding) and negative (threatening) social cues, rather than a simple amplification of threat processing. Importantly, these associations remained significant while accounting for the effects of co‐occurring psychopathological symptoms. In this way, the flattened subjective value gradient we report may contribute to blunted affective discrimination across social contexts—consistent with neurocognitive perspectives indicating compromised integration of social information in lonely individuals [37]. These findings nuance sociocognitive models by highlighting that loneliness may not universally heighten threat reactivity; instead, it may attenuate the relative weighting of social outcomes overall, yielding diminished approach motivation even in contexts where engagement would be beneficial. This interpretation is in agreement with evidence that loneliness can alter neural circuits implicated in trust, reward processing, and social cognition in a manner distinct from social anxiety or generalized threat hypervigilance [38].

Short‐term manipulation of social expectations produced comparatively modest and inconsistent effects on engagement decisions. Although positive social expectancy reliably shifted proximal appraisals, these changes did not translate into robust or replicable changes in choice behavior, nor did they substantially moderate the effects of loneliness. Rather than indicating a failure of the manipulation, this pattern suggests that the valuation processes governing social engagement in loneliness may be relatively stable and resistant to brief expectancy shifts. This dissociation suggests that while social expectations are malleable at the level of appraisal, the valuation processes that govern engagement decisions in loneliness may be relatively stable and resistant to brief contextual perturbations. This pattern contrasts with evidence from reminiscence‐based interventions, which involve repeated and structured engagement with autobiographical social memories and have been shown to reduce loneliness and enhance perceived social connectedness over time [39, 40]. Rather than transiently biasing expectations, such interventions may operate by gradually recalibrating social value representations and social alignment processes through sustained reinforcement of positive social experiences, consistent with predictive and reward‐based accounts of loneliness [35].

Certain limitations should be acknowledged. First, although the study employed repeated, choice‐based social decision tasks, the scenarios were necessarily hypothetical and administered online, limiting generalization to real‐world contexts. Also, as participants were recruited from Polish research panels, replication in more diverse cultural contexts will be important for assessing generalizability. In addition, the vignette attributes were experimentally assigned rather than independently norm‐validated. Although the manipulations produced highly consistent and theoretically coherent behavioral effects across both studies, participants may nevertheless have differed in how they subjectively construed specific social situations. Future work combining independently validated social scenarios, participant‐level ratings of perceived reward, effort, and threat, and ecologically informed approaches such as behavioral observation, experience sampling, or longitudinal designs could help clarify how altered social valuation processes manifest in everyday interactions over time. Second, while the priming manipulation reliably influenced proximal appraisals, its effects on engagement decisions were modest and less consistent across studies, suggesting that brief expectation‐based manipulations may be insufficient to modify the more stable valuation processes associated with loneliness. This finding delineates a boundary condition on short‐term priming rather than ruling out the potential efficacy of longer‐term or repeated interventions. Third, although participants were randomly assigned to priming conditions, one panel showed a modest age imbalance across conditions; importantly, age was included as a covariate in all primary analyses, and controlling for age did not alter the pattern of results. Fourth, the primary outcome was a binary engagement decision, which may not capture more fine‐grained variation in motivational strength; however, this limitation was mitigated by the inclusion of continuous appraisal measures and convergent exploratory analyses that provided insight into underlying valuation processes. Fifth, the two panels differed in several demographic, social, and clinical characteristics; however, these differences were generally small in magnitude. Importantly, the core findings were replicated across studies despite this heterogeneity. This consistency suggests that the principal pattern of altered social valuation was relatively robust across somewhat heterogeneous participant samples. The main exception concerned effort sensitivity, which was observed only in study 2. One possibility is that effort‐related valuation processes may be more context‐ and individual‐dependent than reward‐related valuation, consistent with broader effort‐discounting literature highlighting substantial heterogeneity across effort domains and task structures [41]. Accordingly, the effort‐sensitivity finding should be interpreted cautiously and requires replication in samples designed to directly examine potential demographic and clinical moderators. Finally, loneliness was assessed as an individual difference rather than experimentally manipulated, which limits causal inference about the origins of altered valuation processes, although the repeated‐decision design allows robust inference regarding how loneliness relates to social choice behavior.

These findings offer some implications for interventions targeting loneliness. Approaches that focus primarily on increasing the quantity of social contact—such as encouraging greater participation or expanding social networks—may be insufficient if individuals continue to assign relatively low subjective value to social opportunities. Consistent with meta‐analytic evidence showing greater efficacy of interventions targeting maladaptive social cognition over contact‐based strategies [40, 42], our results suggest that loneliness may be maintained in part by altered social valuation processes rather than lack of opportunity alone. Moreover, a recent systematic review of Theory of Change–based interventions indicates that psychological strategies, including cognitive–behavioral therapy, reminiscence therapy, and social identity interventions, produce moderate reductions in loneliness by targeting the cognitive and emotional processes that sustain social isolation [43]. Accordingly, interventions may be more effective if they explicitly target the recalibration of social value. This perspective aligns with emerging work implicating dysregulated reward processing in loneliness vulnerability, where blunted social reward may perpetuate withdrawal despite available opportunities [44, 45, 46]. Clinically, this implies a shift away from purely exposure‐based approaches toward interventions that tolerate delayed behavioral change and prioritize sustained updating of social value representations. Measures of social value compression may further serve as proximal outcome measures, enabling trials to track mechanistic change in motivational processes that precede observable engagement.

In conclusion, the present study demonstrates that loneliness is associated with systematic alterations in social decision‐making that extend beyond heightened threat vigilance or transient affective states. Across two large preregistered studies, loneliness consistently predicted reduced social engagement and less favorable appraisal of social opportunities, accompanied by diminished sensitivity to social rewards and threats and a compressed subjective value range. Together, these findings support a view of loneliness as a condition characterized by altered social valuation, in which potentially rewarding interactions are insufficiently differentiated from less advantageous alternatives. By integrating behavioral and cognitive approaches with analyses of social valuation, this work provides mechanistic insight into why social withdrawal in loneliness may persist even when opportunities for connection are available. These results highlight social valuation processes as promising targets for future theoretical models and interventions aimed at reducing loneliness and promoting sustained social engagement. Future studies using formal model comparison approaches may help further refine understanding of how different decision components contribute to social engagement behavior.

Author Contributions

Conceptualization—B.M., Methodology—B.M. and W.G., Software—B.M., Formal analysis—B.M., Resources—B.M. and W.G., Data curation—B.M., Writing – original draft—W.G. and B.M., Visualization—B.M., Supervision—B.M., Project administration—W.G., Funding acquisition—B.M.

Conflicts of Interest

B.M. receives honoraria for his associate editor roles at the Journal of Affective Disorders and Psychoneuroendocrinology from Elsevier.

Supporting information

Supplementary Materials: nyas70382‐sup‐0001‐SuppMat.docx