Abstract
Quasi likelihood-based methods for estimating linear and log-linear Poisson Network Autoregression models with $p$ lags and covariates. Tools for testing the linearity versus several non-linear alternatives. Tools for simulation of multivariate count distributions, from linear and non-linear PNAR models, by using a specific copula construction. References include: Armillotta, M. and K. Fokianos (2023). "Nonlinear network autoregression". Annals of Statistics, 51(6): 2526–2552; Armillotta, M. and K. Fokianos (2024). "Count network autoregression". Journal of Time Series Analysis, 45(4): 584–612; Armillotta, M., Tsagris, M. and Fokianos, K. (2024). "Inference for Network Count Time Series with the R Package PNAR". The R Journal, 15/4: 255–269.
Introduction
In fields such as epidemiology and criminology, data often manifests as multivariate count time series—for instance, daily infection rates across cities or monthly crime counts across urban blocks. Analyzing such data requires accounting for both temporal dependencies and the relational network structure between entities. Traditional models often fail to bridge these two dimensions effectively. The PNAR R package addresses this by providing a robust framework for Poisson Network Autoregressive models, enabling estimation, non-linear testing, and simulation of network-dependent count data.
Modeling Framework
The core of the PNAR package lies in its implementation of Quasi-Maximum Likelihood Estimation (QMLE). This method allows for consistent parameter estimation in both linear and log-linear models, even if the data distribution deviates from a pure Poisson process, provided the mean structure is accurately specified.
To capture complex dynamics, the package extends beyond simple linear effects to include several non-linear specifications:
- Smooth Transition (ST-PNAR): Models dynamics that change gradually via a transition function.
- Threshold (T-PNAR): Captures abrupt shifts in dynamics when a network-weighted average exceeds a specific cutoff.
- Intercept Drift (ID-PNAR): Accounts for non-linear shifts in the baseline event rate.
A key innovation in PNAR is treating network effects as weighted averages of a node's neighbors' past values. This approach significantly reduces the parameter space, allowing the simultaneous modeling of hundreds of nodes.
Statistical Inference and Testing
To prevent overfitting and justify the use of complex models, PNAR includes score tests (Lagrange Multiplier tests) to evaluate linearity against non-linear alternatives. A common challenge in threshold modeling is the "nuisance parameter" problem, where the threshold value is only defined under the alternative hypothesis. PNAR automates the resolution of this issue using Davies’ bound and bootstrap methods, providing reliable p-values for researchers.
Datasets and Simulation
The package includes real-world data and synthetic generators to facilitate empirical research and methodological validation.
The Chicago Crime Dataset serves as a primary benchmark, containing monthly burglary counts across 552 census blocks (2010–2015) along with a spatial adjacency matrix (crime_W). For theoretical testing, PNAR provides functions to generate networks based on established models.
Furthermore, the package utilizes a specific copula construction for simulation. This allows users to generate multivariate counts that maintain Poisson marginal distributions while adhering to the dependencies defined by a network adjacency matrix.
Practical Application and Scalability
PNAR is designed for high-dimensional problems. It successfully handles the Chicago dataset's 552 nodes and 72 time points, a scale that often bottlenecks traditional multivariate models. For model selection, the package provides visual tools like lin_ic_plot and log_lin_ic_plot, which use Information Criteria (AIC, BIC, QIC) to help users identify the optimal lag order ($p$).
Conclusion
While PNAR offers powerful tools for network count time series, users should be mindful of computational costs during bootstrap testing for very large networks. The package includes doParallel support to address this. Additionally, while the package defaults to enforcing stationarity for stability, users can disable these constraints if they have a specialized understanding of their network's dynamics. Future developments may expand these tools to handle overdispersed counts or dynamic networks.
Summary updated