Recent years have seen a resurgence of interest in using agent-based models (ABMs) for network formation research, leveraging their analytical flexibility and granularity to capture complex mechanisms that shape empirically observed macroscopic patterns. Many of these mechanisms — such as thresholds, cognitive states, and frames — are difficult or impossible to observe directly. Yet, ABMs provide a means to infer their relative importance in shaping social networks structures.

Despite this potential, integrating ABMs with empirical network analysis remains largely "artisanal", lacking a systematic estimation framework to quantify uncertainty and automatically fit various summary statistics. Additionally, ABMs are often validated against a single observed network, whereas developing robust middle-range theories requires fitting models to multiple networks within a given context.

To address these challenges, we introduce ByPy, a Python-based framework that applies amortized Bayesian inference to fit arbitrarily descriptive ABMs to social network data. ByPy enables researchers to estimate uncertainty around model parameters, assess their role in network formation, and control for classical network processes such as homophily, reciprocity, and transitivity. Furthermore, ByPy incorporates a method for estimating how ABM parameters vary across multiple networks within the same context — analogous to random coefficients in multilevel SAOMs. We demonstrate ByPy’s capabilities by estimating threshold- and emotion-based network formation mechanisms in a single-case advice network and multiple classroom-based friendship networks, respectively.

Pre-Exposure Prophylaxis (PrEP) is a highly effective method for preventing HIV infection, especially for people at higher risk of HIV acquisition. It may not only prevent HIV acquisition for the treated individual but also their sexual partners. We used a trial emulation approach in agent-based models (ABMs) to assess the spillover of PrEP in a sexual network of men who have sex with men (MSM) in Atlanta, Georgia, from 2015 to 2017. We simulated an individually randomized trial stratified by race in each run of the model. We allowed for possible spillover from an agent’s HIV-uninfected partners and defined coverage as the proportion of an agent’s HIV-uninfected partners receiving treatments. We also let the sexual network update at discrete six-month intervals. We estimated spillover in each simulated trial using an inverse probability-weighted estimator adjusted for race and baseline substance use. Then, we averaged over 100 simulations to obtain estimates and corresponding simulation intervals (SI). The estimated spillover effect was protective but smaller in magnitude than the direct effect. Additionally, the estimates were larger for the contrast of 75% versus 25% coverage (Indirect risk ratio (RR) = 0.72, 95% SI = 0.45, 1.02), compared to the contrasts that reflected only a 25% change in PrEP coverage (Indirect RR = 0.81, 95% SI = 0.62, 0.99). This ABM approach can be used to validly assess causal spillover effects from an agent’s sexual partners to improve the delivery of PrEP to MSM and their sexual partners

Sustainable cooperation in contemporary societies is based on prosocial behaviour driven by a mix of rational calculations of expediency and compliance with perceived normative obligations towards others. However, research on prosocial behaviour has rarely addressed the time-varying, context-dependent nature of underlying motives and their macro-level consequences. In particular, ego’s decision to provide alter with costly support may depend on ego’s framing of the relationship as solidary or instrumental. Moreover, ego’s framing of their relationship with alter may vary over time as a macro-micro feedback of certain contextual features, such as the connectivity of the wider social network. Although the empirical literature on advice and support networks is relatively rich, social network research has rarely attempted to identify formation mechanisms based on different motivations, mostly driven by the availability and power of standard statistical models (e.g., ERGMs or SAOMs) which often lack the required flexibility. To test our hypothesis, we analysed the formation of a network of advice and support among independent freelancers sharing a coworking space in Northern Italy. We collected multiplex network data along with various individual characteristics of the subjects and fitted an empirically calibrated Agent-Based Model to a set of macro-level network summary statistics. The model parsimoniously simulates support seekers’ partner selection and potential givers’ decision-making processes, estimating parameters using Approximate Bayesian Computation. Preliminary results show that the observed support and advice networks can be explained by nodes switching between instrumental reciprocity and normatively based solidaristic prosocial behaviour according to threshold-based changes in overall network density.

Statistical physics provides powerful but often abstract models of collective behavior. Agent-based models (ABMs) can be more expressive and realistic, but their integration in empirical analyses remains challenging. In this work, we demonstrate how simulation-based inference (SBI) can leverage statistical physics models — specifically, the Ising model — as interpretable and principled summary statistics to compare the plausibility of multiple ABMs of social behavior. To illustrate this approach, we study the diffusion of a convention in physics (the metric signature, which involves a choice between two equivalent possibilities). We identify the favorite conventions of 2,277 physicists and their co-authorship and co-citation networks. By solving an inverse Ising problem, we reveal the contribution of local mechanisms of coordination (endogenous to a social network) and global mechanisms of coordination (transcending the network) in the adoption of a convention. The Ising model unveils the structure of the underlying coordination game and the relevant social networks. More interestingly, the parameters of the Ising model measured from the data can be used as summary statistics to compare the likelihood of realistic ABMs of preference-formation using simulation-based inference and amortized Bayesian model comparison. We consider a strategic agent model, a model of global cultural transmission, and a model of local cultural transmission by the imitation of peers. The analysis reveals more evidence for the latter. This work shows that statistical physics models can play an intermediate role in simulation-based inference with complex models, and contributes a method for distinguishing endogenous from exogenous collective behavior in multi-layered networks.

The high rates of prosocial and altruistic behavior observed in humans is largely driven by our embeddedness in social networks. One of the key mechanisms through which networks promote cooperation is clustering. Similarly, altruistic giving can generate chains of generalized giving characterized by structural closure. Here we test arguments linking network clustering in altruistic networks to increases in the overall benefits derived by members and decreases in inequality in the overall benefits. We argue that publicly rejecting acts of giving (as opposed to private discarding) will increase the amount given and overall inequality. Similarly, we argue that framing altruistic contributions as a gain in net benefits (as opposed to a loss to the system) will increase the amount given and overall inequality in benefits. Importantly, we also argue that network clustering will moderate these effects by increasing benefits while decreasing inequality. To evaluate our arguments, we collected experimental data from 40 triads. Participants were endowed each round and could give points to others. Points given were doubled and participants could accept one gift per round. We varied whether rejections were made public or if this information was not disclosed, and whether giving was framed as a doubling of value or a loss to the system. Those data were used to train a large-scale simulation varying the clustering of an underlying network of 1,000 agents. Results show that clustering increases giving and decreases inequality while also moderating our manipulations in predictable ways. Implications and future directions are discussed.