Longitudinal inference for sexual contact networks using simulation-based methods. This package implements a universal discrete-time simulator for sexual contact networks and performs approximate Bayesian computation (ABC) inference.

Install uv if you haven't already. Then clone the repository and install dependencies by running:

uv sync --all-groups

To reproduce the results in accompanying publication, run the following commands. All simulations and ABC are orchestrated using the cook task runner. The following command will generate all results required to create the figures.

uv run cook exec -j 4 "default/inference/adjusted/none/*" empirical/inference/default

It will take a while to run: 96 hours CPU clock time on a M4 MacBook Air or 24 hours wall clock time with four processes. You can modify the number of concurrent processes by changing the -j [number of processes] flag.

Launch a Jupyter notebook and navigate to the notebooks directory which contains notebooks in Jupytext markdown format. Right-click any notebook in the Jupyter explorer and select "Open With > Notebook". Executing the notebooks will generate the figures in the manuscript.

• notebooks/illustration.md: Illustration of the network evolution mechanisms (figure 1).
• notebooks/thermodynamic-limit.md: Validation of analytic expressions in the thermodynamic limit against finite-size simulations (figure 2). This notebook might take another hour or so to run.
• notebooks/inference.md: Prior, posterior, and parameter values reported in the literature (figure 3), samples of summary statistics from the prior-predictive and posterior-predictive distributions (figure 4), and evolution of parameter RMSE as a function of between-wave follow-up period (figure 5).

The src/hiv directory contains the Python package responsible for simulation and inference. The main files are:

• algorithm.py: Implementation of nearest-neighbor ABC as a scikit-learn estimator and function to apply regression adjustment to candidate posterior samples.
• simulator.py: Efficient sexual contact network simulator for unipartite graphs.
• util.py: Utility function and implementation of NumpyGraph, a NumPy-based graph representation. In short, nodes and edges are stored as NumPy arrays which allows for performant batch-wise updates, e.g., deleting a set of nodes, adding a set of edges, etc. It is very slow for iterative updates.
• scripts: Collection of command-line interfaces.j
  • generate_data.py: Generate samples from the prior predictive distribution of the sexual contact network model. This is primarily used for constructing the ABC reference table.
  • run_abc.py: Run approximate Bayesian computation for observed data given a reference table.