Contributions
- Quantitative analysis of misinformation cascade dynamics on Facebook, comparing 32 conspiracy-theory pages with 35 science-news pages over five years (2010–2014)
- Demonstration that consumers of scientific and conspiracy theories exhibit similar consumption patterns but significantly different cascade dynamics
- Finding that homogeneity in social networks is the primary driver of content diffusion, with cascades predominantly occurring within homogeneous clusters
- Data-driven percolation model that uses homogeneity and polarization parameters to predict cascade size, successfully validated against both science and conspiracy content, and trolling posts
Method
The analysis uses Facebook data from 67 public pages (32 conspiracy, 35 science news, 2 troll) accessed via the Facebook Graph API. Authors define cascade as a sharing tree: the root is the first user to share a post, with subsequent shares forming branches. They introduce edge homogeneity σᵢⱼ = σᵢσⱼ, where user polarization σ = 2ρ − 1 captures the fraction of likes a user puts on conspiracy-related content.
For spreading dynamics, they model users on a small-world network where each user i has opinion ωᵢ ∈ [0,1] and is exposed to news items j with content ϑⱼ ∈ [0,1]. A user shares news when |ωᵢ − ϑⱼ| ≤ δ (sharing threshold). To capture echo chambers, they model connectivity as a signed network with fraction ϕₕₗ of homogeneous links. This is treated as a branching process; the critical cascade size follows from percolation theory, and they fit parameters using inverse Gaussian distribution of first sharers.
Results
Cascade size and maximum degree follow power-law distributions for all three content types (science: γ=2.21, conspiracy: γ=2.47, trolling: γ=2.44). Maximum cascade sizes were 952 (science), 2,422 (conspiracy), 3,945 (trolling).
For cascade lifetime (time between first and last share): both science and conspiracy show peaks at ~1–2 h and ~20 h. Rapid diffusion occurs quickly (24.4% of science news and 20.8% of conspiracy rumors spread within 2 h). However, lifetime versus cascade size diverges by content: science news reaches peak diffusion at ~200 shares with high lifetime variability afterward; conspiracy rumors show positive correlation between lifetime and size.
Edge homogeneity is predominantly positive (>0) in both content categories, confirming that information circulates mainly within echo chambers. For cascade size vs. homogeneity: in science news, higher homogeneity (0.5–0.8) correlates with larger cascades; in conspiracy, lower homogeneity (~0.25) does. The distributions of total vs. homogeneous sharing paths are nearly identical (Kolmogorov–Smirnov test, p > 0.05), confirming homogeneous paths dominate.
The percolation model with parameters (ϕₕₗ=0.56, r=0.01, δ=0.015) and inverse Gaussian first-sharer distribution reproduces real cascade size and height distributions with high accuracy (CCDF match confirmed in SI).
Notes
This is foundational work in the misinformation cascade literature. The key insight—that homogeneity drives diffusion more than intrinsic content quality—has been widely cited and replicated. The percolation model is elegant but simplified; real Facebook behavior includes algorithmic amplification, tie strength variation, and temporal dynamics the model abstracts away. The comparison of science vs. conspiracy cascades remains novel: both spread via social proof, but assimilation differs (science reaches saturation quickly; conspiracies grow more slowly). Strength: large-scale, multi-year dataset with careful statistical validation. Weakness: observational only (no intervention); causality between homogeneity and cascade size is correlational.