Flexible Modeling of Nonstationary Extremal Dependence Using Spatially-Fused LASSO and Ridge Penalties
Functions for modeling nonstationary extremal dependence using locally-stationary max-stable processes with LASSO and ridge penalization. The provided code is in support of Shao, X., Hazra, A., Richards, J., and Huser, R. (2023+). Flexible modeling of non-stationary extremal dependence using spatially-fused LASSO and ridge penalties. ArXiv.
The two main R scripts are:
- `Modeling.R` - Fits the extremal dependence model with a provided dataset. Can be run for either the simulation study or the application.
- `Summary.R` - Provides a summary of the fitted model with related plots and tables.
Auxillary scripts include:
- `Simulate_data.R` - A simple simulation of Brown-Resnick processes with nonstationary extremal dependence;
- `Algorithm1.R` - Function for Algorithm 1 described in the paper;
- `Merge_subr.R` - Function for the subregion merging process described in the paper;
- `Lambda_tuning.R` - Function for the
$\mathbf{\lambda}$ -tuning described in the paper; - `Fit.R` - Some fitting function for convenience, using r-optim;
- `Objectives.R` - Pairwise likelihood functions for the Brown-Resnick process (and inverted counterpart);
- `Utils.R` - Various other utility functions.
Included in the repo are two Rdata files:
- `NepalExtended.Rdata` - The gridded data of monthly maximum temperature dataset from Nepal and its surrounding Himalayan and sub-Himalayan regions used in the data application of the paper. This file includes the marginal parameter estiamtes (GEV parameters) derived using the Max-and-Smooth method: postman.mu (location), postman.sigma (scale), postman.xi (shape).
- `Simulated.Rdata` - Simulated data from `Simulate_data.R`, including the coordinate and true (dependence) parameter information. The true partition is partition P1 mentioned in the paper.
Some further remarks:
- The current program only works for gridded data, but an extension to general lattice data is available.
- Nonstationarity is assumed for the input data.
- Input data must be renormalized to unit Fréchet margins to fit the max-stable processes.
- Some difficulties in range estimation may emerge.