Integrating behavioral experimental findings into dynamical models to inform social change interventions

1. Problem Statement

The paper addresses a critical disconnect in the literature regarding how to stimulate large-scale adoption of new products or behaviors (e.g., sustainable technologies, health behaviors) to solve global challenges.

• The Gap: Two dominant research traditions exist but rarely integrate:
  1. Individual-level approaches (microeconomics, psychology): Focus on decision-making drivers and utility but often fail to extrapolate to collective dynamics due to non-linear feedback loops.
  2. Computational/Complexity approaches (agent-based modeling, statistical physics): Focus on network propagation and collective dynamics but often rely on simplifying assumptions about individual behavior (e.g., homogeneous thresholds) that lack empirical grounding.
• The Core Challenge: Existing "seeding policies" (strategies to select initial adopters to maximize spread) often rely on network centrality metrics (degree, betweenness) or assume arbitrary threshold distributions. They neglect empirically measured individual-level behavioral drivers, potentially leading to suboptimal interventions.
• Specific Question: Can individual-level thresholds for adoption be derived from discrete choice data and integrated into dynamical models to improve the efficacy of social change interventions?

2. Methodology

The authors propose a framework that bridges micro-level choice theory with macro-level complex contagion models.

A. Theoretical Framework: From Utility to Thresholds

The authors reinterpret the Complex Contagion threshold model through Discrete Choice Theory.

• Utility Decomposition: An individual $n$'s utility for adopting product $i$ ($U_{ni}$) is the sum of attribute utility ($U^{(A)}_{ni}$) and social utility ($U^{(S)}_{ni}$).
  • $U_{ni} = U^{(A)}_{ni} + U^{(S)}_{ni}$
  • $U^{(A)}_{ni} = \sum \beta_{nk} x_{ki}$ (Linear combination of product attributes).
  • $U^{(S)}_{ni} = \gamma_n s_{ni}$ (Linear function of the social signal $s_{ni}$, the fraction of neighbors who adopted).
• Threshold Derivation: The adoption threshold ($\tau_{ni}$) is defined as the minimum social signal required for the utility of adoption to exceed the status-quo utility ($U^{(0)}_n$).
  • Resistance ($R_{ni}$) is defined as the gap between status-quo utility and attribute utility: $R_{ni} = U^{(0)}_n - U^{(A)}_{ni}$.
  • The threshold is derived as:
    $$\tau_{ni} = \frac{R_{ni}}{\gamma_n}$$
• Classification:
  • Susceptible Adopters: $0 < \tau_{ni} < 1$ (Adoption depends on social signal).
  • Innovators: $\tau_{ni} \leq 0$ (Adopt regardless of social signal).
  • Laggards/Independents: $\tau_{ni} \geq 1$ (Never adopt regardless of social signal).

B. Experimental Validation

To estimate these parameters empirically, the authors conducted two Choice-Based Conjoint (CBC) experiments:

1. Policy Support (PS): Support for carbon-capturing energy policies.
2. App Adoption (AA): Adoption of a new instant messaging app.

• Design: Participants chose between hypothetical products defined by specific attributes and varying levels of "social signal" (percentage of friends adopting).
• Estimation: A Hierarchical Bayes (HB) algorithm was used to estimate individual-level parameters ($\beta_{nk}$, $\gamma_n$, $U^{(0)}_n$) from the choice data. These parameters were then used to calculate individual thresholds via the derived formula.
• Validation: The estimated thresholds were tested for out-of-sample predictive power against hold-out tasks.

C. Simulation and Seeding Policies

The authors integrated the estimated thresholds into Agent-Based Simulations (ABS) on real-world social networks (Add Health dataset).

• Goal: Maximize the final number of adopters relative to the cost of intervention.
• Cost Scenarios:
  1. Preference-based cost: Cost depends on the individual's resistance (harder to persuade high-resistance individuals).
  2. Centrality-based cost: Cost depends on the node's network centrality (influencers are expensive).
• Policies Compared:
  • Structure-based: Degree, Betweenness, Closeness, Collective Influence, Uncalibrated Complex Centrality (assumes arbitrary thresholds).
  • Behavior-based: Low-threshold policy, Neighborhood Susceptibility (targets nodes with many low-threshold neighbors), Calibrated Complex Centrality (uses estimated thresholds).

3. Key Contributions

1. Theoretical Integration: The paper provides a mathematical derivation linking individual utility functions (from microeconomics) to the threshold parameters used in complex contagion models. This allows thresholds to be estimated rather than assumed.
2. Empirical Estimation of Thresholds: It demonstrates that individual-level thresholds can be reliably estimated from standard choice experiments, validating their predictive power for individual decisions.
3. Heterogeneity Discovery: The study reveals that threshold distributions are highly product-dependent. For example, energy policy support showed a bimodal distribution (many independent adopters), while app adoption showed a more uniform distribution (more susceptible adopters).
4. Optimization of Seeding Strategies: The paper establishes that the "best" seeding policy is not universal; it depends on the cost structure and the availability of behavioral data.

4. Key Results

• Predictive Validity: Thresholds estimated from experimental data significantly outperformed random thresholds in predicting individual choices in hold-out tasks.
• Context Dependence: The proportion of "independent adopters" (those unaffected by social signals) varied significantly between contexts (higher in policy support, lower in app adoption), highlighting that social influence mechanisms are context-specific.
• Seeding Policy Performance:
  • Preference-based Cost: When the cost to persuade an individual depends on their resistance, the Neighborhood Susceptibility policy (targeting nodes surrounded by susceptible individuals) significantly outperformed all structure-based policies and the uncalibrated complex centrality.
  • Centrality-based Cost: When the cost depends on network centrality (e.g., paying influencers), the Calibrated Complex Centrality (using estimated thresholds) significantly outperformed all other policies.
  • Crucial Finding: The Uncalibrated Complex Centrality (which assumes arbitrary/homogeneous thresholds) failed to perform well, proving that accurate behavioral data is essential for this metric to work.
• Robustness: The advantages of behavior-based policies held even when only a fraction (e.g., 75-90%) of the network was surveyed, though performance degraded with highly incomplete data (50%).

5. Significance and Implications

• Policy Design: The study argues that effective social change interventions cannot rely solely on network topology. Policymakers must measure individual preferences and resistance to design optimal seeding strategies.
• Beyond Complex Contagion: The authors extend their framework to show that Simple Contagion (Bass model) and Preferential Attachment models can also be reinterpreted through individual-level choice theory, suggesting a unified micro-to-macro framework for various diffusion processes.
• Practical Application: The proposed method (conjoint experiments + HB estimation) offers a scalable, data-driven tool for practitioners to identify the most effective "seeds" for campaigns, moving beyond the "influencer" heuristic to a "susceptibility-aware" heuristic.
• Limitations & Future Work: The framework assumes individual preferences are static during the diffusion process. Future work needs to address scenarios where collective outcomes feedback to alter individual preferences (dynamic thresholds).

In summary, this paper successfully bridges the gap between micro-level behavioral economics and macro-level network science, providing a rigorous, data-driven methodology to optimize social interventions by accounting for the heterogeneity of individual decision-making.