A behavioral signature for quantifying the social value of interpersonal relationships with specific others

Humans are highly social, and insufficient social connection is linked to adverse mental and physical health outcomes, motivating the view that social connection is a basic human need. Social stimuli and interactions are intrinsically rewarding, engaging reward-related neural systems, and people will often trade material rewards for social experiences. However, individuals vary in their desired amount and type of social interaction and in their orientation toward outcomes affecting others. This variability has sparked efforts to quantify the value of social phenomena. Prior work has quantified value in social contexts (e.g., information sharing, decisions, single interactions), but relatively little has targeted the value of specific interpersonal relationships. Existing models often treat the social context as a residual or assume constancy across partners, and welfare-tradeoff measures can be sensitive to transient affect and socioeconomic factors, limiting their reliability as relationship-value indices. The present study defines interpersonal social value using concepts of scarcity and opportunity cost, operationalized via how people prioritize time and effort in activities with particular others. The authors develop a behavioral signature of social value—a weighted set of common social activities reflecting an idealized allocation of finite leisure time—and quantify the value of specific partners (parent, friend, acquaintance) by how strongly individuals’ reported activities with each partner express this signature. The approach avoids direct introspective valuation (which can be biased) and focuses on constrained, prioritized behavior.

The authors situate their work within literatures on: (a) intrinsic reward value of social stimuli (neuroimaging evidence of reward-system engagement by social information and interactions); (b) behavioral economic findings showing willingness to forgo monetary rewards for social stimuli; (c) variability in social approach/avoidance and sociability; and (d) models of social valuation and welfare tradeoffs. They highlight gaps: prior quantifications largely address the value of isolated social actions or generic social context, not the value assigned to specific interpersonal relationships. Residual-context terms in prior models do not isolate partner-specific value, and welfare-tradeoff parameters are sensitive to transient affect and socioeconomic background, undermining generalizability. Grounding valuation in opportunity cost/time allocation follows established value-operationalization strategies in related fields, while avoiding confounds from resource availability and introspective biases.

Design: Two-phase study (exploratory and confirmatory) totaling N = 1,111 participants (N_exploratory = 476; N_confirmatory = 635) across multiple independent subsamples (SONA undergraduate pool, MTurk, Prolific). Each participant nominated one parent, one close friend, and one acquaintance. IRB-approved; pre-registered confirmatory phase (08/03/2022); data and code available on OSF.

Exploratory phase procedures:

1. Sourcing activities: Participants listed up to 25 past activities done with another person. Researchers consolidated items to higher-level categories to balance generalizability with specificity. Two activity sets were produced for robustness: SONA-sourced (70 items) and MTurk-sourced (56 items).

2. Deriving activity value weights via Maximum Difference (MaxDiff) scaling: Independent samples completed MaxDiff tasks where, under a free-time, no-obligation scenario, participants chose the most and least likely activity to do from 4-item sets (53 or 42 sets depending on pool). MaxDiff administered separately for SONA- and MTurk-derived activity pools.

Modeling: Hierarchical Bayesian logistic regression (brms) with overparameterized dummy coding estimated item coefficients (log-odds) representing the likelihood of selecting an activity as most vs least likely. Coefficients were random across participants; fixed-effect posterior means were scaled by random-effect variances to account for individual preference variability. Weakly informative priors; Bernoulli likelihood; 8 chains, 1,000 iterations (500 warmup), target accept = .95, step-size .05, max treedepth 15. Resulting coefficients are activity weights constituting the behavioral signature (ordinally rankable due to zero-sum MaxDiff). Sensitivity analyses indicated robustness to weight scaling and high-variance-item inclusion (correlations of alternative scoring variants r = 0.81–0.98).

3. Computing social value scores: For each nominated partner (parent, friend, acquaintance), participants provided likelihood ratings: How likely they would engage in each signature activity with that partner under average conditions and easy access. Social value for a partner = dot product of that partner’s likelihood vector and the activity-weight vector (behavioral signature), yielding three scores per participant (one per partner).

Validation measures (exploratory): Relationship quality (IPPA), social loss aversion (upset if unable to spend time with partner), actual and ideal time spent with each partner, one-shot dictator games (monetary allocation preferences across partner pairs), and a forced-choice item on whom to spend a free afternoon with.

Inferential approach: Bayesian estimation; posterior summaries with means and 89% HDIs; ROPE = [-0.1, 0.1]; evidence judged robust if HDI excluded 0 or lay outside ROPE, moderate if part of HDI outside ROPE.

Confirmatory phase procedures:

• Larger samples (four subsamples) recruited from MTurk and Prolific; MTurk data screened for fraud and supplemented with Prolific data due to high fraud rates. Fully crossed design between activity-weight source (SONA vs MTurk) and validation sample (MTurk vs Prolific). Final N: MTurk=181 (82 SONA-weights; 99 MTurk-weights); Prolific=454 (233 SONA-weights; 221 MTurk-weights).

• Measures replicated from exploratory phase (relationship quality, social loss aversion), with time-spent questions dropped to reduce burden.

• Added multi-trial social decision-making tasks: Repeated choices allocating money or time between partner pairs (delay-discounting format), allowing estimation of choice preferences across varying reward/timing contexts.

• Added affiliative behavior preferences: Forced-choice items across six affiliative behaviors (advice seeking, celebrating, sharing positive or negative news, lending money, having dinner), crossed with all partner pairs (18 items). Bayesian logistic regression with standard normal priors used for choice models.

Post hoc psychometrics (confirmatory samples with largest N):

• Reliability of likelihood ratings via bifactor models: ω_total, ω_hierarchical, and ECV computed (psych::omega).
• Measurement invariance (configural) of activity weights across partners (RMSEA acceptable ~0.07–0.08; CFI poor 0.42–0.53), suggesting practical unidimensionality within each partner set but failing strict invariance tests.
• Shuffled-weights test: Randomly permuted activity-weight mappings, recomputed scores, and reran key correlations and pairwise contrasts; shuffled distributions centered near null.
• Incremental validity over unit-weighted likelihood means: Bayesian multiple regressions including both social value scores and mean likelihood scores predicting relationship quality, social loss aversion, dictator-game preferences, and time-spent forced choice.
• Discriminant validity: Cross-partner correlations of social value with non-matching partners’ relationship quality and social loss aversion; partial correlations controlling matched-partner social value.
• Predictive validity controlling for other relationship facets: Regressed partner-pair choice preferences on differences in social value, relationship quality, and social loss aversion (standardized predictors; Bayesian regression).

Sample and signature: Total N = 1,111 across 10 subsamples; two independently sourced activity sets (SONA: 70 items; MTurk: 56 items) with MaxDiff-derived hierarchical Bayesian weights formed the behavioral signature.

Exploratory results:

• Relationship quality correlations: Friends mean r ≈ 0.31 (evidence in 4/5 subsamples); acquaintances mean r ≈ 0.19 (evidence in 3/5); parents mean r ≈ -0.04 (unstable across subsamples).
• Social loss aversion correlations: Generally positive for all partners (mean r ≈ 0.11; evidence in 3/5 subsamples for each).
• Time-spent associations: Inconsistent and small; largely not reducible to overall time. Examples: parents actual time mean r ≈ 0.11 (robust in 2/5); friends actual time mean r ≈ -0.17 (robust in 2/5); acquaintances ideal time mean r ≈ 0.14 (evidence in 5/5).
• Decision preferences (dictator game; forced choice time): Coefficients generally in predicted directions (favoring partners with higher social value), but with large posterior variances, likely due to smaller Ns and model complexity.
• Pairwise differences (sanity checks; Table 3): Social value differences typically favored closer over more distant partners (e.g., Friend > Acquaintance; Parent > Acquaintance) with positive mean differences and HDIs excluding 0 in many subsamples.

Confirmatory results (replication):

• Relationship quality (Table 5): Social value correlated positively across all partners and subsamples: parent mean r ≈ 0.17 (robust 4/4), friend mean r ≈ 0.26 (robust 4/4), acquaintance mean r ≈ 0.24 (robust 4/4).
• Social loss aversion (Table 5): Positive associations: parent mean r ≈ 0.19 (robust 4/4), friend mean r ≈ 0.21 (robust 4/4), acquaintance mean r ≈ 0.24 (evidence 3/4).
• Decision preferences (one-shot): Multiple regressions showed social value differences predicted dictator-game allocations in all partner pairings; forced-choice time-spent effects were in the expected directions but less consistent (especially friend vs acquaintance, with higher variance).

Confirmatory results (expansion):

• Multi-trial social decision-making: Robust evidence that social value scores predict choices across monetary and social outcomes for all partner pairings (parent vs friend; parent vs acquaintance; friend vs acquaintance). Effects remained when adjusting for relationship quality. Model-implied probabilities showed clear preference shifts with ±1.5 SD manipulations of partner-specific social value.
• Affiliative behaviors: Social value reliably predicted preferences across six affiliative behaviors, most consistently for parent vs friend comparisons; generally expected directions for parent vs acquaintance and friend vs acquaintance with somewhat less consistency.

Psychometrics and validity:

• Reliability of likelihood ratings: High internal consistency and strong general-factor saturation. • SONA-weights sample: ω_total ≈ 0.982–0.989; ω_hierarchical ≈ 0.732–0.883; ECV ≈ 0.567–0.741 (across partners). • MTurk-weights sample: ω_total ≈ 0.972–0.985; ω_hierarchical ≈ 0.651–0.799; ECV ≈ 0.484–0.703.
• Shuffled-weights test: Correlations and pairwise differences centered near zero when weight-activity mappings were randomized (e.g., r_shuffled around -0.15 to 0.15 SD ≈ 0.10–0.14), supporting the meaningful contribution of the behavioral weights.
• Incremental validity: Social value scores remained associated with relationship quality, social loss aversion, and dictator-game preferences after controlling for unit-weighted likelihood means; the forced-choice time-spent association was largely accounted for by the likelihood component.
• Discriminant validity: Most cross-partner correlations ≤ 0.12 threshold; two exceptions (parent social value with friend/acquaintance relationship quality) attenuated notably after partialing out matched-partner social value (partial r ≈ 0.198 and 0.077), indicating shared individual-difference variance.

Overall: The behavioral-signature-based social value scores showed convergent validity with relationship quality and social loss aversion, predicted social decision preferences in one-shot and multi-trial paradigms, demonstrated good reliability of input ratings, and were not reducible to total time spent with partners.

The study addresses the question of how to quantify the value individuals assign to specific interpersonal relationships by operationalizing value through opportunity costs in time/effort allocation. By engineering a behavioral signature of social value (MaxDiff-weighted activities) and measuring its expression in partner-specific likelihoods, the authors show that social value scores track with key relational constructs (relationship quality, social loss aversion) and with choices in social decision tasks. This provides mechanistic specificity beyond traditional models emphasizing the value of isolated actions, demonstrating that partner-specific valuation is a distinct and behaviorally relevant construct. Importantly, associations with actual or ideal time spent were small and inconsistent, indicating that the measure captures prioritized, constrained behavior rather than raw contact frequency. Post hoc analyses suggest that while social value contributes uniquely, relationship quality and social loss aversion often exert stronger predictive influence for the behavioral outcomes studied. The approach offers a framework for partitioning action value into components attributable to the partner and to situational features, with potential applications in refining theories of value-based decision-making and relationship dynamics, and informing research on disruptions of value processes (e.g., in psychopathology).

The authors introduce and validate a practical method to quantify interpersonal relationship value grounded in opportunity cost and scarcity: a behavioral signature of social value derived from MaxDiff-weighted activities and expressed via partner-specific likelihood ratings. Across large, independent samples, the resulting social value scores correlate with relationship quality and social loss aversion and predict social decision preferences in both one-shot and multi-trial paradigms. The measure is not reducible to overall time spent and exhibits favorable psychometric characteristics. Future work should generalize signatures across cultures, incorporate personalized or quasi-personalized weights, integrate activity effort costs and goal relevance, and test neural correlates and real-world ecological validity (e.g., via EMA).

Key limitations include: (1) Need to establish ecological validity by linking scores to real-world daily social behavior (e.g., via EMA). (2) Lack of neural validation; future studies should test correlations with value-related brain signals when evaluating specific partners. (3) Cultural generalizability is uncertain; items/weights may reflect WEIRD populations, and norms around activities/partners may constrain choices in other cultures, necessitating culturally adapted items. (4) Measurement invariance across partners was only partially supported (acceptable RMSEA but poor CFI), warranting further psychometric refinement. (5) Weights are group-derived; personalized or quasi-personalized weights could improve precision, especially for high-variance activities. (6) Potential conflation with activity effort; future designs should incorporate effort estimates or prompts controlling for effort. (7) Conceptual bidirectionality and contextual influences (relationship dynamics affect activities and vice versa) not fully modeled. (8) Some behavioral predictions were modest and, in several cases, were better accounted for by relationship quality or social loss aversion, indicating overlapping but distinct constructs.