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

TL;DR

This work tackles the gap between micro-level decision making and macro-level diffusion by deriving individual adoption thresholds from choice-based conjoint experiments and embedding those thresholds into threshold-based spreading models. Using two CBC studies (policy support and app adoption) and Hierarchical Bayes estimation, the authors obtain individual utilities and social-signal weights, from which thresholds $\tau_{ni}$ are computed and validated via out-of-sample predictive accuracy. The simulations show that threshold-aware seeding policies can outperform traditional structure-based or naive strategies, depending on the cost structure, and that the approach generalizes to other diffusion frameworks such as the Bass model and preferential attachment. Overall, the paper provides a data-driven framework to calibrate social spreading with empirical human drivers, enhancing the design of policies and interventions for large-scale behavioral change.

Abstract

Addressing global challenges -- from public health to climate change -- often involves stimulating the large-scale adoption of new products or behaviors. Research traditions that focus on individual decision making suggest that achieving this objective requires better identifying the drivers of individual adoption choices. On the other hand, computational approaches rooted in complexity science focus on maximizing the propagation of a given product or behavior throughout social networks of interconnected adopters. The integration of these two perspectives -- although advocated by several research communities -- has remained elusive so far. Here we show how achieving this integration could inform seeding policies to facilitate the large-scale adoption of a given behavior or product. Drawing on complex contagion and discrete choice theories, we propose a method to estimate individual-level thresholds to adoption, and validate its predictive power in two choice experiments. By integrating the estimated thresholds into computational simulations, we show that state-of-the-art seeding methods for social influence maximization might be suboptimal if they neglect individual-level behavioral drivers, which can be corrected through the proposed experimental method.

Figures (16)

• Figure 1: Estimating individual–level thresholds.(A) The complex contagion theory assumes that individual–level adoption choices are determined by the decision-makers' threshold. On the other hand, individual-level perspectives focus on estimating individuals' utilities of adopting from choice data. Our work reconciles the two perspectives by reinterpreting the individual-level thresholds in terms of individuals' attribute and social utilities. (B) For a susceptible adopter, the status-quo utility is initially larger than the utility from adopting. As the number of adopters increases, so does her social utility. The threshold is defined as the minimal level of social signal at which the utility from adopting exceeds the utility from not adopting. (C) For both experiments, the individual-level thresholds estimated from experimental data hold out-of-sample predictive power, as illustrated by their superior accuracy compared to a random-threshold baseline. (D) In general, different products exhibit different threshold distributions, as illustrated by the two examples provided here (instant messaging app in blue; energy policy in orange). (E) The distribution of the proportion of independent adopters (as opposed to susceptible adopters) is significantly lower for the app adoption experiment (AA, in blue) than for the policy support experiment (PS, in orange), which highlights the importance of context for the distribution of individual thresholds. (F) An individual is susceptible for adopting a given product when her resistance is positive and lower than the marginal utility of social signal, which corresponds to the gray stripe in the $\gamma-R$ diagram. There are significantly more observations that fall within the gray stripe in the AA experiment than in the PS experiment, which explains the higher percentage of susceptible adopters in the AA experiment. Data in this panel is based on a sample of products, as described in Supplementary Note \ref{['secSI:product_sampling']}.
• Figure 2: Relative performance of seeding policies.(A) Nodes selected by different policies in an illustrative network structure: The highest-degree node (in blue) has the largest number of connections, but the highest-neighborhood susceptibility node (in dark red) has the largest number of connections to low-threshold nodes (in red). (B, C) Relative performance of seeding policies under a preference–based cost structure, measured through the average rank defined in the main text, for the policy support experiment and the app adoption experiment, respectively. The neighborhood susceptibility policy based on the estimated thresholds significantly outperforms the other policies. (D, E) Relative performance of seeding policies under a centrality–based cost structure for the policy support experiment and the app adoption experiment, respectively. The complex centrality policy based on the estimated thresholds significantly outperforms the other policies.
• Figure S1: Average utilities across respondents. The y axis contains all attributes and levels. The x axis shows the average utility of an attribute level together with the 95 $\%$ confidence interval. On average, the percentage of friends endorsing the policy (labeled as percentage_adopters) has a positive partworth utility.
• Figure S2: Average threshold over individuals (y axis) for each policy (x axis). The error bars represent the 95$\%$ confidence intervals for the mean. The policies are sorted by the average threshold. We observe that some policies have a lower average threshold, which implies that they are more attractive to the respondents.
• Figure S3: Average utilities across respondents. The y axis contains all attributes and levels. The x axis shows the average utility together with the 95 $\%$ confidence interval. We can see that on average, the percentage of adopters has a large, positive utility.