Effective Bayesian Modeling of Large Spatiotemporal Count Data Using Autoregressive Gamma Processes
Yifan Cheng, Cheng Li · 2025
Read Effective Bayesian Modeling of Large Spatiotemporal Count Data Using Autoregressive Gamma Processes by Yifan Cheng, Cheng Li (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-19915.txt.
Introduction
Yifan Cheng and Cheng Li offer a new way to model counts that shift across place and time in a 2025 paper. Such counts arise in health, crime, and sales data, and old models often force slow or rough fits. Their method builds the hidden rates from a chain of gamma steps that stays steady over time.
The split form lets a sampler draw each part in turn, with no rough steps in the loop. A sparse map of near places keeps the cost in line with the data size. Tests on made-up data and real sets show sound fits and good forecasts for new places and times.
Catalog record
| CREATOR | Yifan Cheng, Cheng Li |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 153 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-19915
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Effective Bayesian Modeling of Large Spatiotemporal Count Data Using Autoregressive Gamma Processes Yifan Cheng, Cheng Li Abstract We put forward a new Bayesian modeling strategy for spatiotemporal count data that enables efficient posterior sampling. Most previous models for such data decompose logarithms of the response Poisson rates into fixed effects and spatial random effects, where the latter is typically assumed to follow a latent Gaussian process, the conditional autoregressive model, or the intrinsic conditional autoregressive model. Since log-Gaussian is not conjugate to Poisson, such implementations must resort to either approximation methods like INLA or Metropolis moves on latent states in MCMC algorithms for model fitting and exhibit several approximation and posterior sampling challenges. Instead of modeling logarithms of spatiotemporal frailties jointly as a Gaussian process, we construct a spatiotemporal autoregressive gamma process guaranteed stationary across the time dimension. We decompose latent Poisson variables to permit fully conjugate Gibbs sampling of spatiotemporal frailties and design a sparse spatial dependence structure to get a linear computational complexity that facilitates efficient posterior computation. Our model permits convenient Bayesian predictive machinery based on posterior samples that delivers satisfactory performance in predicting at new spatial locations and time intervals. We have performed extensive simulation experiments and real data analyses, which corroborated our model’s accurate parameter estimation, model fitting, and out-of-sample prediction capabilities. Keywords: spatiotemporal count data, Gibbs sampler, autoregressive gamma process, Vecchia approximation. 1 Introduction Spatial and spatiotemporal count data are ubiquitous in many scientific fields, such as ecology ( Estes et al. 2018 ) , epidemiology ( Ward et al. 2023 ) , and transcriptomics ( Zhao et al. 2021 ) . Bayesian models for such data enable flexible hierarchical structures, convenient posterior sampling, and robust uncertainty quantification. A classic statistical model for spatial and spatiotemporal count data is the latent Gaussian model (LGP, Diggle et al. 1998 ; Christensen and Waagepetersen 2004 ), where the count data is assumed to follow a standard generalized linear model such as the log-linear model, whose log-transformed parameters are specified as a linear combination of fixed effects and spatial random effects. For point-referenced spatial data, the spatial random effects are typically assumed to follow a continuous latent Gaussian process, characterizing the spatial or spatiotemporal dependence and leading to a well-defined stochastic process over continuous domains. For areal data, the spatial random effects are typically assumed to follow the conditional autoregressive model (CAR) or the intrinsic conditional autoregressive model (ICAR, Besag 1974 ; Mardia 1988 ; Besag et al. 1991 ; Besag and Kooperberg 1995 ). Computationally, both LGP and CAR models are often fit by either the integrated nested Laplace approximation (INLA, Rue et al. 2009 ; Blangiardo et al. 2013 ) or Markov chain Monte Carlo (MCMC) algorithms that require Metropolis moves for latent states ( Lee et al. 2018 ) . For example, a recent survey on the spatial and spatiotemporal analysis of COVID-19 epidemiology reveals that almost all statistical models adopted in practice belong to either the LGP or ICAR class of models and used either INLA or MCMC for model fitting ( Nazia et al. 2022 ) . While the LGP and CAR models have achieved considerable empirical success in statistical applications, there exist several potential drawbacks in current statistical models and computational strategies. First, modeling the spatial random effects as Gaussian processes implies that the rate parameter of a Poisson response variable follows a log-Gaussian process, which is not conjugate to the Poisson distribution. Therefore, to compute the posterior distribution, one has to rely on either approximation methods such as INLA or Metropolis moves on latent states in an MCMC algorithm. INLA typically depends on tessellation of the spatial domain and incurs an approximation error that does not necessarily vanish with large samples. For MCMC algorithms, the dimension of latent states gets large as the number of spatial locations becomes large, and thus, the Metropolis moves may result in slow mixing of posterior chains and low effective sample sizes. Second, the log-Gaussian process is known to lead to potentially large gradients of log-likelihoods with respect to the latent states and model parameters in the covariance functions ( Christensen and Waagepetersen 2004 ) , which makes exploration of the full posterior difficult and requires numerical truncation. Third, realizations of an LGP model often demonstrate the whitening effect ( Morales-Navarrete et al. 2024 ) , due to implicit restrictions on the correlation structure caused by the nugget for spatial count data ( De Oliveira 2013 ) . The literature offers many alternative strategies to overcome the aforementioned modeling and computational challenges. For instance, De Oliveira 2013 ; De Oliveira 2014 studied several classes of transformed covariance functions to remove discontinuity in the covariance function caused by the nugget and reduce the whitening effect. More recently, new efficient MCMC algorithms based on the Pólya-Gamma sampler have been proposed in D’Angelo and Canale 2023 for the Poisson model and in Bansal et al. 2021 for spatial negative binomial model. Bradley et al. 2018 and Bradley and Clinch 2025 introduced new classes of multivariate log-gamma distributions as a fully conjugate prior of the Poisson model for spatial count data, which results in a fully conjugate posterior that does not require MCMC sampling. This work focuses on Bayesian modeling of spatiotemporal count data and proposes a new modeling strategy that enables efficient and fully conjugate Gibbs sampling of spatiotemporal frailties and model parameters. Instead of assuming a log-Gaussian process on the spatial random effects as in most of the previous literature, we take the Poisson frailties associated with all spatial locations corresponding to each time interval as a vector and generalize scalar autoregressive gamma processes for time series ( Gourieroux and Jasiak 2006 , Creal 2017 ) to model the spatiotemporal frailties, thus extending prior applications of autoregressive gamma processes from the time series setting and the time-to-event network setting ( Zhai et al. 2025 ) to the spatiotemporal setting. We make four main contributions, as follows. First, we establish a spatiotemporal autoregressive gamma process provably stationary across the time dimension. Different from LGP and CAR, our construction does not involve any Gaussian latent variables. Second, we adopt a novel decomposition of latent Poisson variables that allows fully conjugate Gibbs sampling of the frailties and scalar parameters c,\,\kappa,\,\rho in Section 2 . Third, we design a spatial dependence structure built on a sparse neighborhood graph that leads to a computational complexity linear in both the number of spatial locations and the number of time intervals in each posterior sampling iteration, thus facilitating efficient posterior computation. Fourth, we develop a complete set of Bayesian predictive machinery for new points in both spatial and temporal dimensions, based on posterior samples of spatiotemporal frailties and model parameters. The rest of this paper is organized as follows. Section 2 introduces our Bayesian spatiotemporal count model with prior specification and a rigorous theory on the stationarity condition for the spatiotemporal frailty process. Section 3 details the posterior sampling steps for all model parameters. Section 4 elaborates on procedures for predicting at future time intervals and unseen spatial locations. Section 5 conducts extensive simulation experiments that justify our model’s accurate parameter estimation, model fitting, and prediction capabilities on par with alternative methods. Section 6