EnviroStat-package:
Statistical analysis of environmental space-time processes
Description
EnviroStat provides functions for spatio-temporal modeling of
environmental processes and designing monitoring networks for
them based on an approach described in Le and Zidek (2006).
Details
The modeling approach offered by this package has a number of features:
- Conditional on knowing the process parameters the
environmental process is assumed to have (after a suitable
transformation if necessary) to be a Gaussian random field
(GRF).
At every spatial location, the process can yield a
multiplicity of random responses such as air pollutant
concentrations.
The approach used in the package lies within a Bayesian
hierarchical modeling framework However for computational
expediency empirical shortcuts are made at higher levels of the
hierarchical setup. Thus for example most hyperparameters are
fitted using a type II maximum likelihood approach, eliminating
the need for the the user to specify them. Thus the package can
handle large fields of monitoring networks, say with 600 or more
spatial sites.
The approach does not assume a stationary GRF. Instead it
takes a nonparametric approach where the spatial covariance
matrix is left completely unspecified and instead endowed with a
prior distribution with a hypercovariance matrix that can be
modeled at level two of the hierarchy, making the method quite
robust against non-stationarity in the random field.
It presents a approach for designing monitoring networks
based on the well-known warping method of Sampson and Guttorp
(1992) as developed with Wendy Meiring.
It allows for missing data, providing that these data are
missing in blocs of time, which after a regional trend is
fitted, then become exchangeable. For then the blocs of
residuals can be permuted the get a decreasing or increasing
staircase pattern in the data matrix something that is required
in the approach.
It has been empirically assessed in a number of major
applications and found to yield well calibrated prediction
intervals. For example, a 95% interval will cover their
predictands about 95% of the time.
References
Le, Nhu D. and James V. Zidek. Statistical Analysis of
Environmental Space-Time Processes. Springer, New York, 2006.
See Also
See the package vignette for a guided example of complete
analysis using the package and the manual for details of
individual functions.