Copula-based models for spatially dependent cylindrical data

Technical Summary: Copula-based Models for Spatially Dependent Cylindrical Data

Problem Statement
Cylindrical data, consisting of paired circular (angle) and linear (intensity) variables, are prevalent in fields such as meteorology, oceanography, and ecology. A primary challenge in analyzing these data is jointly modeling the angular–linear dependence and spatial autocorrelation while maintaining flexibility in marginal specifications. Existing literature often relies on copula-based constructions but typically treats copula parameters as constants unrelated to covariates and restricts marginal regression to linear predictors, thereby ignoring spatial correlation. Furthermore, standard Gaussian-based approaches often fail to accommodate non-Gaussian margins or non-linear dependencies between the circular and linear components.

Methodology
The authors propose a Structured Additive Conditional Copula Regression Model for cylindrical data, termed the Partially Wrapped Copula (PWC) model. The methodology is built on the following components:

• Partially Wrapped Approach: Instead of defining the joint distribution directly on the cylinder $S^1 \times \mathbb{R}$, the model adopts a "wrapped" approach. The circular response $\phi_1(s)$ is treated as the modulo $2\pi$ of an underlying latent linear spatial process $Y_1(s)$. A latent winding number process $k(s) \in \mathbb{Z}$ is introduced such that $Y_1(s) = \phi_1(s) + 2\pi k(s)$, allowing the model to operate on the unwrapped linear space $\mathbb{R}^2$ before wrapping the circular component.
• Conditional Copula Framework: Leveraging Sklar's theorem, the joint conditional distribution is decoupled into marginal distributions and a copula function that governs dependence. The copula parameter $\rho$ is allowed to vary as a function of covariates and space, addressing the gap in existing cylindrical models where copula parameters are often static.
• Marginal Distributions:
  • The circular component is modeled using a Wrapped Gaussian (WG) distribution, induced by an underlying Gaussian Process (GP).
  • The linear component follows a Log-Normal distribution, also driven by an underlying GP.
• Spatial Dependence and Non-Stationarity: Spatial effects are modeled using Gaussian Markov Random Fields (GMRFs) derived from Stochastic Partial Differential Equations (SPDEs). This approach approximates Gaussian Random Fields (GRFs) to ensure computational efficiency via sparse precision matrices. Crucially, the model allows for non-stationarity in both the mean and covariance structures by letting the SPDE parameters (range and variance) vary smoothly over space via log-linear predictors.
• Bayesian Inference: Posterior estimation is performed via Markov Chain Monte Carlo (MCMC) simulation. The authors employ a two-stage approach: first, marginal models are fitted (sampling latent winding numbers and GP parameters), and second, the copula regression coefficients are estimated conditional on the fitted marginals. This strategy reduces computational burden compared to joint estimation while maintaining valid uncertainty quantification.

Key Contributions

1. Covariate-Dependent Copulas for Cylindrical Data: The paper extends the application of covariate-dependent copulas to cylindrical settings, allowing the dependence structure between wind direction and speed (or similar variables) to vary based on environmental covariates.
2. Flexible Marginal and Spatial Modeling: By combining wrapped Gaussian processes with distributional regression, the model accommodates non-Gaussian margins and non-linear spatial effects.
3. Non-Stationary Spatial Processes: The framework explicitly models non-stationary spatial dependence in both the mean and covariance of the latent processes, utilizing the SPDE-GMRF framework to handle the computational complexity of large spatial datasets.
4. Latent Winding Number Recovery: The introduction of a latent winding number process facilitates the use of standard linear spatial modeling techniques for circular data, avoiding the complexities of partial periodicity in the joint likelihood.

Results

• Simulation Study: A simulation study evaluated the model's ability to identify tail dependence (Gaussian, Clayton, and Gumbel copulas) and covariate effects. The results indicated that information criteria (DIC and WAIC) were generally effective in selecting the correct copula family. Asymmetric copulas (Clayton and Gumbel) were more easily distinguished from the Gaussian copula due to their distinct tail properties, particularly as sample sizes increased.
• Application to German Wind Data: The model was applied to hourly wind direction and speed data from 289 weather stations in Germany during a storm event (January 2025).
  • Model Selection: The Gumbel copula with a covariate-dependent parameter provided the best fit, outperforming models with constant parameters or independence assumptions. This indicated the presence of upper-tail dependence between wind direction and speed during the storm.
  • Covariate Effects: The analysis revealed that higher air pressure was associated with lower wind speeds and altered directions, while air temperature and humidity increased wind speeds. The dependence parameter $\rho$ was positively associated with air pressure and visibility, suggesting stronger dependence under stable atmospheric conditions.
  • Spatial Prediction: The model successfully captured wind dynamics in northern and central Germany. Predictive performance was slightly reduced in the southern mountainous regions (Alps), attributed to complex orographic effects and the current mesh resolution. The posterior mean of the copula parameter showed a spatial gradient, with stronger dependence in the north compared to the south.

Significance and Claims
The authors claim that the proposed PWC model offers a flexible framework for analyzing spatially dependent cylindrical data that overcomes the limitations of existing approaches. By integrating covariate effects into the dependence structure and allowing for non-stationary spatial processes, the model provides a more accurate representation of complex environmental phenomena like wind behavior. The paper emphasizes that while the application focuses on wind, the framework is generalizable to any spatial circular–linear data. The work highlights the importance of moving beyond static dependence assumptions and linear predictors to capture the nuanced relationships between direction and intensity in spatially correlated datasets. Future work is suggested to extend the framework to fully spatio-temporal settings and explore more flexible marginal distributions (e.g., wrapped skew-normal) and semi-parametric copula formulations.