R Package iglm: Regression under Interference in Connected Populations
Scale regression to networks with provable guarantees and custom model terms
Cornelius Fritz and Michael Schweinberger introduced the R package iglm, a comprehensive framework for regression under interference in connected populations. This package enables the study of spillover and other phenomena, where outcomes in one unit are affected by predictors in connected units. It addresses a key limitation of existing packages by offering scalability and provable theoretical guarantees.
On the computational side, iglm relies on scalable methods applicable to small and large datasets, solving a convex optimization program based on pseudo-likelihoods using Minorization-Maximization and Quasi-Newton algorithms. Statistically, the framework comes with provable theoretical guarantees. To increase versatility, users can add custom-built model terms. The package was showcased on two datasets: hate speech on X and communications among students.
- iglm handles spillover effects in connected populations with provable theoretical guarantees
- Scalable convex optimization via pseudo-likelihoods, Minorization-Maximization, and Quasi-Newton algorithms
- Tested on hate speech from X and student communications, supports custom model terms
Why It Matters
Enables robust analysis of network interference, critical for social media and epidemiology studies