Disentangling material, social, and cognitive determinants of human behavior and beliefs

Human behavior and beliefs in social interactions are shaped by a complex interplay of material payoffs, social influences (descriptive and injunctive norms; authority messaging), and cognitive processes (cognitive dissonance, social projection, logic constraints, errors). The relative importance of these drivers has been debated across philosophical traditions and modeled in economics, social psychology, and network science. This study addresses the research question: What are the relative weights of material, social, and cognitive determinants in human decision-making and in the dynamics of personal norms and beliefs about others? The authors integrate these factors in a mathematical model where actions depend on expected material payoffs, cognitive dissonance, conformity to peers, expected disapproval (normative expectations), and compliance with authority. Belief dynamics are explicitly modeled via cognitive dissonance, social projection, logic constraints, learning from observed behavior, and authority influence. A long-term behavioral experiment is conducted to validate and parameterize the model, tracking coevolution of actions, personal norms, and empirical and normative expectations, with and without authority messaging about the group-optimal action.

Model: Individuals choose a continuous action x. Each has a personal norm y (appropriate action), an empirical expectation x̄ (expected average peer action), and a normative expectation ȳ (expected peer view of appropriate action). An authority promotes G. Utility function: u = A1·π(x, x̄) − (1/2)A2(x − y)² − (1/2)A3(x − x̄)² − (1/2)A4(x − G)², capturing material payoffs, cognitive dissonance, conformity with peers’ actions, and compliance with authority. Best response: x = max(0, B0·θ + B1·y + B2·ȳ + B3·x̄ + B4·G) summarized as x = max(0, B0 + By + Bŷ + Bx̄ + BG), with weights Bi related to Ai (ΣBi=1). Belief dynamics (per period): y′ = y + α1(x − y) + β1(X − y) + γ1(G − y); ȳ′ = ȳ + α2(y − ȳ) + β2(X − ȳ) + γ2(G − ȳ); x̄′ = x̄ + α3(ȳ − x̄) + β3(X − x̄) + γ3(G − x̄). Here X is the observed average action of current groupmates. Terms capture cognitive dissonance (align attitudes with past action), conformity and learning from peers (move beliefs toward observed X), social projection and logic constraints, and compliance with authority messaging. All parameters are individual-specific. Experiment: 35 daily rounds (spring 2021), online Common Pool Resources (CPR) game. Each round, subjects are randomly re-grouped (n=6), receive 30-point endowment, and choose investment x into a “Common Account” (CPR extraction) vs a safe “Personal Account.” Group investment Z = Σxi yields collective return P(Z) = bZ − 0.5 d Z² with b=15, d=1/6; individual share vi = xi/Z; payoff πi = π0 + vi·P − c·xi (c is extraction cost; payoff calculator provided). Nash equilibrium per-person x_NE = 24; group-optimal x_opt = 14. Two treatments: without messaging; with authority messaging each round: “Please note that the total group profit is maximized if each player contributes 14 points to the Common Account.” Elicitation: Before choosing x each round, subjects report: personal norm y (“should” invest), empirical expectation x̄ (others’ investments; averaged), and normative expectation ȳ (what others think the subject should invest; averaged). Belief elicitation is incentivized by accuracy. After each round, subjects see their payoffs and groupmates’ actions. Additional measures: Big Five, risk preferences, rule compliance (Ball task), demographics, and Social Value Orientation (SVO). Estimation strategy: Mean group estimation (Pesaran–Smith) applied to individual-level regressions for best response (x on θ, y, ȳ, x̄ with intercept) and belief dynamics (parameters αi, βi, γi as applicable). For actions, per subject estimate: B0, B1, B2, B3, intercept C, and error σ (B4 not estimable since G constant=14, absorbed by intercept). For beliefs, estimate αi, βi (and γi with messaging), intercept, and error. Model space: for actions, 32 combinations (presence/absence of predictors); for beliefs, 8 (no messaging) or 16 (with messaging). Address multicollinearity via condition numbers/variance decompositions; apply ridge regression when needed. Model selection via AICc; use model averaging with AIC weights; confidence intervals via nonparametric bootstrap. Validate via simulations using estimated parameters: predicted next-step trajectories and multi-round simulations (500 runs) with random reshuffling, initialized at observed round-1 values. Cluster analyses: k-means and Gaussian mixtures on estimated parameters; and Dynamic Time Warping clusters on action time-series; stratified analyses by SVO and rule-following.

- Behavior levels: Mean extraction converged around x ≈ 22 (below Nash equilibrium x_NE=24; above social optimum x_opt=14). Subjects invest more than they believe is right (x > y) and more than they expect others to invest (x > x̄). Personal norms equilibrate faster than other variables. - Messaging effects on means/variances: Messaging did not change average extraction or payoffs but increased their variance. Messaging reduced y, ȳ, and x̄ toward G=14 and reduced variance of these beliefs. - Best-response weights (mean estimates): Without messaging: B1 (personal norms/cognitive dissonance) = 0.32 (largest); B3 (conformity to expected peers’ actions) = 0.16; B0 (material payoffs) = 0.11; B2 (normative expectations) = 0.09. With messaging: B1 drops to 0.19; B3 rises to 0.21; B0 ≈ 0.12; B2 ≈ 0.10. Thus, messaging decreases the weight of personal norms and increases the weight of conformity; material and normative factors remain smaller and largely unchanged. - Belief dynamics: Cognitive forces (α1 cognitive dissonance, α2 social projection, α3 logic constraints) and learning from peers (βi) are comparable in magnitude for y and ȳ; for x̄, learning from peers (β3) dominates. Authority influence (γi) is strongest for personal norms (γ1 > 3× combined cognitive and conformity effects) and substantial for normative expectations (≈1.5× combined effects); weak for empirical expectations (x̄), as peers’ actions are observed directly. Messaging reduces αi and reduces β1 and β2 on y and ȳ. - Heterogeneity and clusters: Strong between-individual variation (skewed parameter distributions). Utility-parameter clusters (k-means): Cluster 1 (norm-driven; highest B1) invests least; Cluster 2 (conformity-driven; highest B3) invests most; Cluster 3 (balanced) shows comparable B0–B3. Belief-dynamics clusters: without messaging, types include stable-norms (low α1, β1), peer-driven (high βi), and mixed (αi and βi comparable); with messaging, a messaging-dominated cluster (high γ1, γ2) and a mixed cluster. - Type differences: Prosocial vs individualistic (SVO): With messaging, prosocials reduce extraction while individualists increase it. Prosocials have higher B1 and lower B2, B3; are more responsive to messaging (γ). Rule-followers vs rule-breakers: With messaging, rule-followers decrease, rule-breakers increase extraction; rule-followers show higher B1 and greater γ; messaging lowers B1 and raises B3 for both, with the largest increase in B3 among rule-breakers. These offsetting responses explain the null average effect of messaging on x but increased variance. - Gender: Small differences overall; without messaging, males weigh material payoffs more, females weigh personal norms and normative expectations more; males more affected by authority messaging. - Model validation: Predicted and simulated mean trajectories align well with observed dynamics, with some mismatch for x and x̄ under messaging. Subjects lost up to ~15–50% of potential payoffs due to non-material influences.

The study quantifies the relative contributions of material, social, and cognitive factors to decision-making and belief dynamics in a CPR setting. Findings show that personal norms (cognitive dissonance) and conformity to expected peers’ actions dominate decision weights, while material payoffs and normative expectations are less influential. Authority messaging shifts decision determinants by reducing reliance on personal norms and increasing conformity, and it strongly shapes personal norms and normative expectations themselves. Belief changes arise from both cognitive processes and social learning, underscoring that understanding behavioral dynamics requires modeling the coevolution of personal norms and beliefs about others. Heterogeneity is central: distinct behavioral types respond differently to messaging (e.g., prosocials and rule-followers versus individualists and rule-breakers), producing canceling effects on averages but larger variances. These insights clarify why interventions may show limited average impact yet meaningfully reshape belief landscapes and subgroup behaviors, with implications for targeting policies and anticipating population responses to shocks or nudges.

This work integrates material, social, and cognitive determinants into a single dynamic framework, validated by a 35-day CPR experiment with incentivized belief elicitation. It directly estimates the weights of these factors in actions and belief updates, revealing the primacy of personal norms and conformity, the modest role of material payoffs and normative expectations, and the substantial influence of authority messaging on beliefs. Contributions include: (1) an empirically parameterized model of coevolving actions and beliefs; (2) quantified decision weights and belief-update parameters; (3) evidence of strong heterogeneity and identifiable behavioral types; and (4) demonstration that messaging effects depend on individual types and can increase behavioral dispersion without changing means. Future research should test generalizability across non-WEIRD populations and varied cultural contexts; examine other social interactions (coordination, public goods, collective risk, sharing, rule-following); explore messaging leveraging social identity and norms; and develop richer models of learning, utility (including asymmetric/discontinuous forms), and discrete actions, enabling personalized or cluster-targeted interventions.

- External validity: Participants were WEIRD and the context was CPR/resource extraction; generalizability to other populations and interaction types is untested. - Experimental design: Random reshuffling may have lowered the salience of normative expectations; duration (35 days) may have been insufficient for full convergence of some variables. - Measurement and priming: Belief and norm elicitation occurred before choices and was incentivized; no control arm without elicitation, raising concerns about priming effects (though prior evidence is mixed). - Parameter identification: Authority level G was constant (14), preventing estimation of B4 (authority weight in best response), which was absorbed by the intercept. For some subjects, limited variation in θ hindered B0 estimation, potentially biasing material payoff weight downward. - Model fit: A subset of subjects (e.g., conditional cooperators/compliers) and individuals with stable norms (α1≈β1≈0) were not well captured, contributing to mismatches (especially for x and x̄ under messaging).