Abstract
We consider network autoregressive models for count data with a non-random neighborhood structure. The main methodological contribution is the development of conditions that guarantee stability and valid statistical inference for such models. We consider both cases of fixed and increasing network dimension and we show that quasi-likelihood inference provides consistent and asymptotically normally distributed estimators. The work is complemented by simulation results and a data example.
Introduction
In many modern applications, researchers track discrete counts over time across connected entities, such as flu cases in different cities or social media interactions across a network. These datasets represent multivariate count time series influenced by a specific network structure. Traditional tools like Vector Autoregression (VAR) are typically designed for continuous data, and existing count-based models often struggle with the "curse of dimensionality" as the number of nodes ($N$) grows.
Armillotta and Fokianos (2024) address this gap by providing a rigorous statistical framework for Count Network Autoregression (PNAR). Their work establishes the mathematical conditions for stability and valid statistical inference in regimes where both the time series length ($T$) and the network size ($N$) increase—a critical requirement for analyzing modern large-scale networks.
Methodology
The authors introduce two primary variants of the Poisson Network Autoregression (PNAR) model:
- Linear PNAR: The expected count for a node is modeled as a weighted sum of its own past (the "momentum effect") and the average past counts of its neighbors (the "network effect").
- Log-linear PNAR: This variant uses a logarithmic link function, allowing for negative coefficients and the seamless integration of external covariates such as demographics or environmental factors.
Technical Innovations
To ensure the model remains stable and does not explode to infinity, the authors demonstrate that the system is stationary if the sum of the momentum and network effect coefficients is less than 1. For estimation, they employ Quasi-Likelihood Inference (QMLE). This approach is computationally efficient as it treats nodes as independent in the "working" likelihood, while a "sandwich" covariance estimator is used to provide valid confidence intervals that account for actual correlations between neighbors.
Furthermore, the authors utilize $L^p$-Near Epoch Dependence (NED) theory to handle the infinite-dimensional nature of growing networks. This sophisticated mathematical tool proves that nodes far apart in the network eventually become independent enough for the Law of Large Numbers to hold. Additionally, a copula-based construction is used to model joint dependence without compromising the tractability of the marginal models.
Empirical Results
Simulation and Robustness
The PNAR framework was tested using Stochastic Block Models (SBM) and Erdős-Rényi graphs. Simulations confirmed that as $N$ and $T$ increase, parameter estimates converge to true values and follow a Gaussian distribution. Notably, the model proved robust to minor misspecifications in the network structure; estimators remain consistent even if some links are incorrectly identified, provided the total error remains bounded.
Case Study: Chicago Burglary Data
The model's real-world utility was demonstrated using monthly burglary counts across 552 census block groups in South Chicago (2010–2015). Using geographic adjacency to define the network, a PNAR(2) model—incorporating a two-month lag—was identified as the optimal fit via the Bayesian Information Criterion (BIC).
The PNAR model achieved a Root Mean Square Error (RMSE) of 0.038, significantly outperforming the standard Spatio-Temporal Autoregressive Moving Average (STARMA) model, which recorded an RMSE of 0.079. This result highlights the model's superior ability to capture the "rippling" effect of crime through connected neighborhoods.
Conclusion and Future Work
While QMLE is computationally efficient, it is not the most statistically efficient method. The authors suggest that a two-step Generalized Estimating Equations (GEE) approach could further refine precision. Future research directions include extending the theory to time-varying networks, where the structure evolves over time, and exploring high-dimensional covariate selection within the network framework.
Summary updated