The aim of this paper is to propose
three approaches to analyse spatio- temporal environmental data supposing
that the above mentioned data can be easily represented using dynamic space-time
interaction models. From the current literature on this topic, appear that
the study of a space-time dynamic phenomenon, can be performed in the data
domain, using space-time auto-correlation or state- space approach
and in the frequencies domain, through bi-dimensional spectral analysis.
Departing from this consideration
we consider:
1) Generalized Spatio-Temporal Autoregressive
Moving Average Models (GSTARMA) when the data
are stationary
both in time and space.
2) State- Space Spatio Temporal Dynamic
Model (SSSTD) when the data are not stationary in time or
and in space
and equally spaced.
3) Bi-dimensional Spectral Analysis (BSA) when are unknown the dynamic mechanisms of the system.
The three proposed methodologies have been compared with an application to environmental data on sea-water pollution observed in the Adriatic sea.
MAURO COLI
Dipartimento di Metodi Quantitativi
e Teoria Economica
University G. d'Annunzio
Viale Pindaro, 42
Pescara, 65127, Italy
coli@dmqte.unich.it
Specification Errors in Spatial
Autoregressive Models: Impacts on Parameter Estimation
Francesco Lagona
Gaussian Autoregressive models play
a prominent role in the modeling of spatial lattice series. Important
applications include the analysis of the impacts of environment quality
on risks for human health, production of goods and economic activity, to
mention a few. In these cases, regional frequencies of interest are often
modeled by a Binomial or Poisson distribution depending on a number of
parameters and, according to the empirical bayesian paradigm, a suitable
transformation of the aforementioned parameters is modeled with a spatial
Autoregressive (or Auto-Gaussian) model depending on interaction iperparameters.
Although these hierarchical models are highly parametric, they often
show a good performance, accounting for unequal spatial variation and spatial
dependence of data.
Spatial Autoregressive models can
be viewed as a particular case of the general linear model specification,
with their main feature being that the covariance error matrix is modeled
by some neighborhood structure among observation sites.
Unfortunately, though, in most application
positing the correct correlational structure among errors is more the exception
than the rule, as a specification error in the error covariance matrix
may occur. Two of the most common cases are: (1) the neighborhood structure
may be misspecified (e.g., links that exist are not included and/or links
that do not exist have been included: neighborhood structure specification
error, NSSE), or, more generally, (2) the dependence structure may be misspecified
(e.g., spatial stationariety is misspecified for a non-stationary situation:
correlational structure specification error, CSSE).
This communication is devoted to
study the distributional properties of maximum likelihood estimators under
NSSE or CSSE (e.g. bias, consistency, efficiency), and the relating distortion
of confidence intervals. After reviewing some known asymptotic results,
the main focus will be on the finite lattice properties of these estimators
under various specification errors, with discussion of analytical results
and simulations studies.
FRANCESCO LAGONA
Centre for Economics of Institutions
University "Rome 3"
via Corrado Segre 4
Rome 00146, Italy
lagona@uniroma3.it
Exploring and Modeling Multivariate
Spatial data
Giovanna Jona Lasinio
In the analysis of multivariate spatial
data several difficulties arise. Often, because of the nature itself of
the phenomena under study, we cannot assume independence between observations,
and instead seek to model the specific type of dependence we observe.
We usually model these datasets as random realizations of a multivariate
spatial random field (MSRF), with specific assumptions on the dependencies
among the variables -- which comprise the combined observation vector and
the spatial coordinates. However, before any serious attempt to specify
a model can be made, we have to explore the available data as deeply as
possible. This is made very difficult by the high dimensionality
of the data. Standard multivariate exploratory techniques such as principal
components analysis (PCA) and multidimensional scaling, usually fail
to give us a good representation, as they do not explicitly take into account
the spatial arrangement of observations.
Here an exploratory technique
based on the diagonalization of cross-variogram matrices is described.
Through the definition of a model for the analysis and simulation
of multivariate spatial data, a test procedure for the assumption
of isotropy of multivariate spatial data is proposed. Applications to simulated
and real data will be briefly reported.
GIOVANNA JONA LASINIO
Dipartimento di Statistica, Probabilita
e Stat. Appl.
University of Rome "La Sapienza"
P.le Aldo More 5
Rome 00185, Italy
jona@pow2.sta.uniroma1.it