Royal Statistical Society Environmental Statistics Study Group

Study Group Chair and Secretary

The Environmental Statistics Study Group of the RSS was formed in 1996 under the auspices of the General Applications Section.  Since then ithe committee has been organising meetings on a wide variety of topics of interest to statisticians and other environmental scientists.  Our objectives include acting as a national focus for statisticians involved in environmental work  Similar groups have been formed elsewhere including the American Statistical Association's Section on Statistics and the Environment ( ASA ).  The Study Group also provides a cadre of experts on whom the Society can call for professional advice.

Study Group meetings 2000/2001

Next meeting

Thursday 10 May, 2pm at the RSS (Tea 3pm).

Ecological modelling

MIKE FASHAM (Southampton Oceanography Centre)
Modelling open ocean ecosystems with simple deterministic models

The problem in modelling open ocean ecosystems is that available datasets are usually sparse both in temporal coverage and in the coverage of some of the key ecological variables such as zooplankton. Examples will be given in which simple ecosystem models are fitted to time series using nonlinear optimisation techniques.

SIMON WOOD (University of St. Andrews)
Statistical modelling of population dynamics using mechanism based population dynamic models.

The models most appropriate for mechanistically describing the dynamics of ecological populations are usually non-linear: often to an extent that is inconvenient for statistical work. This talk will cover some attempts to use ecological dynamic models to statistically compare alternative mechanistic models of biological populations.
 
 

Previous meetings

15th Feb 2.00pm, Maths and Stats Dept, Fylde College, Lancaster
University

Tea 3:15
 
 

The topic is   Countryside Survey 2000

 Speakers are

 Colin Barr      (Centre for Ecology and Hydrology, Merlewood)
         Brief review of CS2000 methodology and Examples of results

 Ralph Clarke     (Centre for Ecology and Hydrology, Dorset)
         Statistical issues

 
The Countryside Surveys are unique in generating national statistics on an integrated suite of rural features. Recent changes in the sampling strategy, and shifts in the detail of field methodology, have generated more reliable country-based estimates of features which have both policy relevance and science interest, without losing backward comparability. This talk will present details of the methodology, changes since earlier surveys, and give some examples of results from the CS2000
survey.

Monday 26 March, 2pm at the RSS (Tea 3pm).

Environmental sampling

JIM ZIDEK (University of British Columbia)
Approaches to the design of monitoring networks

With increasing concern about environmental risk, the demand for reliable information has grown rapidly.  This requires efficient statistical methods for analysing available data and observation networks designed to provide timely, reliable data in amounts adequate for the application of those methods. Various approaches to environmental network design will be reviewed.

STEVE THOMPSON (Pennsylvania State University)
Adaptive sampling in surveys, experiments, and observational studies

With adaptive sampling designs, sample selection depends on values observed during the survey allowing, for example, additional sites to be added near hot spots. Adaptive strategies may provide considerably more precise estimates than conventional designs and also allow effective data inference for observational studies with little control over sample selection.

RON SMITH (CEH Edinburgh)
Sampling for estimation of air pollution levels and air pollution environmental effects

Estimating air pollution effects on, for example, vegetation requires sampling of both air pollution levels and vegetation effects. Co-located sampling is probably optimal but not desirable for predicting effects over large areas. The accuracy of model predictions for regions becomes important and determining a suitable monitoring strategy can be critical.
 

October 19th 2000:  GM food and crops

Venue:  RSS Headquarters, Errol St, London.

The GMHT Farm Scale Evaluations
Joe N. Perry (IACR - Rothamsted Experimental Station) & Peter Rothery
(Centre for Ecology & Hydrology, Monks Wood)

The Government is currently sponsoring an experiment, through DETR, called the Farm Scale Evaluations (FSE), to study the effect of the management of four herbicide tolerant genetically modified (GMHT) crops on the biodiversity of farmland wildlife.  The experiment began in 1999; it will end in 2002.  Although the design is simple, the experiment raises a
number of issues, mainly concerned with: (i) the power of ecological studies based on discrete count data; (ii) the size and position of the experimental units; (iii) representativeness of results and their applicability to a wider population of farms; (iv) pragmatic problems of implementation of experiments that are of strong public interest and that attract popular protest.  These issues will be discussed from the viewpoint of those contracted to do the research.
 
 

Risk Assessment and GM Crops

John Fenlon (Horticulture Reserach International)

Environmental impacts appear to be the major concern relating to GM crop release. How far can these impacts be assessed with currently available information, and how can statistical modelling help? The talk will consider the major components of risk using a framework model, and discuss the progress made in calibrating them. In particular I will discuss how
models from other areas of science can be useful.
 
 

Extrapolating beyond experimental boundaries in GM crop assessment

G R Squire, B Marshall, G Ramsay, C Thompson, JW McNicol
Scottish Crop Research Institute, Invergowrie, Dundee DD2 5DA
 

Questions on GM risk are usually referenced to a particular spatial or temporal scale – how far will genes move between fields; how many years will GM organisms persist?  However, the knowledge that has to be used to answer such questions is usually gained through measurements at a smaller scale, on a less complex system, than those of the question itself.  Biological experiments, by definition, impose tight boundaries that restrict the exchange of organisms or genes and the conditions under which the exchanges take place. Risk assessment generally needs to predict what would happen if the boundaries were relaxed or extended, and in doing this, can employ several devices. It should make particular use of biological features that are conservative across organisms and scales; indeed, in many instances of risk assessment, data collected on non-GM organisms can provide the vital breakthroughs in understanding. Moreover, some boundaries can be extended by increasing the range of environmental factors which the organisms experience, without affecting the exchange of organisms with the wider environment. And if a wider ecological risk is minimal, boundaries can be made deliberately leaky so as to allow limited exchanges of organisms that need to be part of the system under study. These points are illustrated by three topical features of GM risk. One is the widespread development of oilseed rape as a feral weed, which was clearly predictable from extant wider knowledge, and would have been more readily demonstrable by widening the scope of measurements and the test environments. The second is gene flow distance, often underestimated in initial studies, but largely predictable in general terms from a basic understanding of pollen vectors and regional scale influences.  The third is the potential occurrence of herbicide tolerant GM ‘superweeds’, which (present knowledge suggests) might be an inconvenience to farmers but not a serious environmental problem. Prior understanding of biological processes provides the key to successful extrapolation in all these instances. More generally, biological science will be better able to answer such questions if it were to find ways of moving more surely between different experimental boundaries, and scales of measurement, rather than continuing to measure ever more accurately at a particular spatial or temporal scale.
 
 
 

22 November at 2pm:  Noise Pollution
 

Venue:  Seminar room 3, Dice Building, Nottingham Trent University
 
 

Noise and aging influences on hearing function: how big is  the problem and how varied are the effects we observe
Adrian Davis, MRC Institute of Hearing Research

Our current research has shown that if a population sample  is taken, social noise exposure from all sources has  increased substantially over the last decade, and is now  probably a potentially greater hazard than occupational  noise in the UK. This increase in risk has not necessarily  led to an increase in hearing thresholds as experienced in a  cross sectional study. Longitudinal studies are needed to  monitor hearing thresholds and hearing function. However,  there is a threefold increase in tinnitus reported by the  group with more exposure. This is a substantial increase  that needs to be replicated.
 
 
 
 

Incidence and attitude to noise
 Philip Wright, Building Research Establishment

BRE has been conducting two national studies of  environmentalnoise on behalf of the Department of the  Environment, Transport and the Regions. The first is concerned  with the incidence of noise, in terms of measured decibel levels;  and the second with people's attitudes tonoise.  This talk will  discuss the design of the two studies, and compare their initial  findings with those of similar studies conducted in 1990 and  1991.
 

Tea at 3.15pm
 
 
 
 

Study Group Meetings 1999/2000

A series of case studies will be discussed at the meeting on the 6th July, with a view to the data being made available for analysis. Results will then be reported at a meeting in the following sessions. Brief descriptions of the data and the scientific issues are provided at case.htm
 
 

April 27 2000, RSS, Errol St, London

Quantitative issues facing the marine environment

Programme

10.00-10.30 Registration and coffee.

10.40 - 11.25 Jacquie McGlade, NERC

Development of Expert Systems for Environmental Management

11.25 - 12.10 Liz Clarke, University of St Andrews

Modelling fish populations in space and time

12.10 - 12.55 Julian Metcalfe, CEFAS

Problems with patterns: from individual behaviour to population models

13.00 -14.00 Lunch

14.00- 14.45 Alistair Murray, British Antarctic Survey

Quantifying spatial scale in the marine environment - some examples from the Southern Ocean.

14.45 - 15.30 Bob Clarke, PML

Measuring biodiversity and structural redundancy in marine communities

15.30-16.10 Tea \

16.15 - 16.50 Peter Challenor, Southampton Oceanography Centre

Some statistical aspects of data assimilation.

Abstracts

Modelling Fish Populations in Space and Time

Liz Clarke, Research Centre for Wildlife Assessment, University of St Andrews

The successful management of fisheries requires accurate estimates of stock size and trends. These estimates are usually obtained using a simple population dynamics model to combine commercial catch data with relative abundance indices obtained from research surveys, thus creating absolute abundance indices. The abundance indices and catch data are commonly available only at the level of the whole stock. Here we apply generalized additive models to spatially referenced research t r awl data to create maps of the spatial distribution of North Sea herring. The outputs of these models are aggregated over the area of interest to provide the input for a Bayesian population dynamics model for the stock. We compare the results from this mo d el with output from the frequentist method currently used for the management of North Sea herring. We discuss the pros and cons of these two methods and their constituent parts and consider some technical issues involved. The next development would be to combine the two stages of the estimation procedure into one model describing both the spatial distribution and population dynamics of the fish: work towards this goal is discussed.

Problems with patterns: from individual behaviour to population models

Julian Metcalfe The Centre for Environment, Fisheries and Aquaculture Science, Lowestoft

Migration is a common feature of the life histories of most of the economically important species of fish. For many, these movements are extensive, and the regular seasonal changes in population distribution that result often have significant impacts on commercial fisheries. Consequently, understanding migration is important to the cost-effective and sustainable exploitation of fish stocks. At Lowestoft, our aim in studying fish migration is to gain a sound understanding of the basic biological (behaviour and physiology) and environmental (water currents, temperature etc.) processes which affect migrations. Our approach has been to use various telemetry techniques t o monitor the behaviour of individual fish in the open sea. In the 70s and 80s we used ship-borne sonar to track individual fish and to learn about their behaviour and movements. This knowledge allowed us to construct " first generation" behaviour-based sim ulation models of the spatial movements of fish populations. However, tracking is severely limited because only one fish can be followed at a time, each fish can only be tracked for a limited period, and sea-going research programmes are very expensive. More recently, we have been using electronic data storage (archival) tags to provided detailed information about patterns of behaviour of many fish for extended (1 year + in some cases) periods. This paper will describe our work with plaice, what we have learnt about its behaviour, and how we have used our knowledge to develop models of population movement. The paper will develop to focus on how we are using complex patterns of behaviour to parameterise our " second generation" models.

Quantifying spatial scale in the marine environment - some examples from the Southern Ocean

Alistair Murray, British Antarctic Survey, Cambridge.

The spatial and temporal scales at which physical and biological processes occur in the ocean define the nature of interactions between organisms and the environment and among the components of the marine biota. The talk will explore what is meant by "scale" and will outline why it is scientifically interesting. Scale is important in designing surveys, analysing data, and buildi n g ecological models. Methodologies for quantifying aspects of scale will be discussed. Examples of spatial analysis of physical and plankton data using spectral analysis and geostatistics will illustrate some of the questions which can readily be addres sed given suitable data sets. The implications of these findings for the design of marine surveys will be explored. Some of the key challenges raised by adopting a scale-based approach will be aired.

Measuring biodiversity and structural redundancy in marine communities

Bob Clarke, Plymouth Marine Laboratory

Biodiversity issues are of major international concern and some of the fundamental problems are statistical. For example, how do we best define and measure biodiversity, in a way that utilises historical data to monitor change? Indices based on taxonomic relatedness of species, with less problematic sampling properties than those of simple species richness, are applied to historic data on in fauna of marine sediments and groundfish surveys aro u nd the UK. Also considered is the issue of structural 'redundancy' in a species assemblage: can one find minimal-sized subsets of species whose multivariate pattern matches that of the community as a whole, and are there several such mutually exclusive subsets (high redundancy)?
 
 

Some statistical aspects of data assimilation

Peter Challenor, Southampton Oceanography Centre, Southampton

The use of numerical modelling in oceanography is increasing all the time. Although today's models are very good at s imulating the ocean there is a need with both physical and biological models to combine the models with data. This combination is known as data assimilation. There are two main reasons for doing this: one is to use the models for forecasting and the secon d is to use the data to estimate parameters in very complex numerical models. The former application is clearly related to weather forecasting and most of the techniques oceanographers use have come from the meteorological community. These techniques includ e the Kalman filter and various approximations to it plus more heuristic methods. Assimilation as estimation is a less well explored technique although a number of papers have addressed this problem particularly with biological models. In this talk I desc ribe a new method of attacking this problem using Bayes Theorem and MCMC techniques. This method will be illustrated with examples from both physical and biological oceanography.
 
 
 
 
 
 

Feb 10:  Gales and Floods at RSS, London at 2pm
 

Speakers

Dick Tabony, Met Office
The lowest possible temperature in the UK

Nick Cook, Bristol University
Extreme wind speeds for the design of structures

Duncan Reed, Institute of Hydrology
Regional Flood Frequency:  What, why and how.
 

Abstracts
 
 

THE LOWEST POSSIBLE TEMPERATURE IN THE UK.

The nuclear industry is obliged to ensure that its equipment is designed to withstand climatological extremes which have a
1:10,000 chance of occurring in a given year.  Estimates of events with this degree of rarity involve recognizing the existence of a physical
limit.

Annual extremes of temperature were compiled for 60 stations from 1914 to 1993.  Stations on coasts and in the Scottish glens displayed an
approach to a lower limit but extreme value plots for stations in the English midlands gave the appearance of being unbounded below.

A combination of physical reasoning and statistics were used to explain this behaviour and a model was developed which estimated
return values of minimum temperature for any location in the UK.
 

Extreme wind speeds for design of structures
 

For many years, extreme wind speeds, and sometimes extreme dynamic pressures, have been fitted to Fisher Tippett Type I distributions
to derive values for use in engineering design.  Recently, there has been a movement towards adopting Generalised Extreme Value (Type III)
distributions, either directly or extrapolated from the Generalised Pareto distribution to account for "observed" departures from Type I.
The Type III distribution predicts an upper limit to the values. With wind speeds, the upper limit corresponds to values (around 50 - 55 m/s
for gusts) for which there is no physical limiting mechansim.  Both Type I and Type III are asymptotic distributions which apply in the
limit as the population from which the extremes are drawn approaches infinity.  The presentation will consider the rate of convergence of
the exact distribution of extremes to the asymptotic model and will show that the annual population of independent wind speed events is
insufficient to achieve convergence.  A method of pre-conditioning the data for fastest convergence to Type I will be demonstrated that
produces results comparable to the Type III on un-conditioned data, but without the implied upper limit.  It is suggested that this
pre-conditioning method is generally applicable whenever the Weibull distribution is a good model for the parent.
 
 

Regional Flood Frequency Analysis: What, Why and How?

The Flood Estimation Handbook (FEH) presents new general procedures for rainfall and flood frequency estimation throughout the
UK. The talk will attempt to:

1.  introduce the FEH statistical procedure for (river) flood  frequency  estimation,
2.  discuss the philosophy behind use of regional frequency analysis methods in hydrology,
3.  illustrate the tools being provided to engineers and hydrologists.

Tea at 3.45pm
 
 
 

Climate change: past and present

December 9th, 3-6pm, LT2, Roger Stevens Building, University of Leeds

======================================================================
The use of Generalised Linear Models in interpreting climate variability

Richard Chandler
Department of Statistical Science
University College London

Abstract:

Analysts of climate data are often interested in quantifying the
nature and extent of any changes which may be taking place. Equally
important is the ability to assess honestly our uncertainty
regarding such changes. Generalised Linear Models
provide an extremely powerful (if underused) framework within which
to study these problems. This talk will focus on some of the issues
which arise when trying to make sense of historical climate records,
and will discuss ways in which these may be addressed
within the GLM framework. Examples from case studies will be used
to illustrate the ideas.

======================================================
Indicators of Climate Change in the UK

Tim Sparks

NERC Institute of Terrestrial Ecology, Monks Wood, UK

In June of this year the Department of Environment, Transport and
Regions (DETR) published a document listing 34 indicators including
examples of climatology, hydrology, socio-economics, agriculture and
the natural environment. The UK government has now commissioned
several "Indicator" projects to cover different aspects of the UK
economy and environment. Readers of RSS News may recall criticism
levelled at some earlier indicators. This talk covers the remit given
by DETR for climate indicators, the process of identifying appropriate
time-series and examples of variables that made it to the publication
stage. The presenter apologies in advance for a bias towards the
natural environment!

======================================================
 

Statistical considerations in climate reconstruction from proxy data

Robert G. Aykroyd (University of Leeds)

To understand recent climate trends and possible future change it is
necessary to  examine the nature of past climatic variability. It is
well-established practice to reconstruct past climate parameters from
proxy variables using a transfer function derived from training data.
Historical proxy measurements can then be used to extend climate
records beyond the limited period of direct instrumental measurements.

Traditional regression-based techniques and an alternative
non-parametric Bayesian approach will be discussed. The effect of low
proxy-climate correlation on accuracy and precision will be examined,
in particular showing that error intervals often span the total
observed range. Finally, a changing pattern in correlation over the
last 100 years will be described which is partly attributed to change
in the length of the growing season.

=========================================================
 
 
 
 

Reducing Uncertainties

October 21st, 2-5pm, RSS Headquarters, Errol St.

Tony O'Hagan, University of Sheffield

"Reducing Uncertainty in Mechanistic Models".

There are several sources of uncertainty that need to be recognised in the use of mechanistic computer models of environmental processes.
The talk will begin by reviewing these, but will focus on two particular sources of uncertainty.  Often, there are model inputs
whose values are not known exactly, but we can learn about these if there are observations of the process itself.  We present a Bayesian
method for calibrating models in this way that allows for remaining uncertainty in the fitted inputs.  A more complex source of error is
deficiencies in the model itself.  We show how the residual errors after calibration can be used to estimate and reduce model
inadequacy.

Howard Grubb, University of Reading

"Uncertainty ingeophysical and meteorological modelling"

Further meetings will be posted as organised.
 

Study Group Meetings 1998/1999

Model Uncertainties

October 22nd, 2-5pm, RSS headquarters, Errol St

Model Uncertainty and uncertainty modelling

Marian Scott, University of Glasgow

Abstract
Sensitivity and Uncertainty Analysis (SA and UA) are modelling tools which have widespread use in fields as diverse as economics, physics, environmental sciences, medicine and social sciences.
Sensitivity analysis is a general methodology used to evaluate the sensitivity of model output to changes in model input, i.e. the rate of change of the response function relative to the input parameters.  It is also closely linked to Uncertainty Analysis, where concern shifts to the evaluation of the uncertainty on the model repsonse as a result of uncertainties on the model unput parameters (parametric uncertainty) and on the model form itself (structural uncertainty).  UA provides a quantification of the overall uncertainty associated with the response.
In this talk, the use of such tools and the challenges presented in their application will be discussed.

Equifinality as a basis for uncertainty estimation in environmental modelling.

Keith Beven, University of Lancaster

Abstract
Experience suggests that because of the inherent limitations of environmental models and the data on which they are based, there are usually many model structures and many parameter sets within model structures that can be considered as acceptable representations of a system of interest.  This equifinality or non-uniqueness requires that the concept of an optimum model be rejected and that techniques be developed to estimate the resulting predictive uncertainty that explicitly recognises the multiple acceptable models, which are often found to be scattered widely in the parameter space.  One such method is the Bayesian Monte Carlo based Generalised Likelihood Uncertainty Estimation (GLUE) approach in which one or more likelihood measures derived from a calibration data set are used to reflect (or modify a prior) belief in the relative probability of different models (and parameter sets) producing acceptable predictions.  These relative probabilities may then be used to weight the predictions associated with each model as the basis for construction of prediction limits.  The approach raises some interesting questions about the nature of physical parameterisations and hypothesis testing in model rejection.

Scenario and paramteric uncertainty in nuclear waste disposal risk assessment: the case of GESAMAC.

David Draper, University of Bath

Abstract
In this talk, I will discuss quantification of uncertainty surrounding the consequences of nuclear waste disposal in deep underground multi-barrier storage facilities, focussing on how uncertainty in scenarios (such as stream flow rates) may be correctly propogated.  My case study will be GESAMAC, a three year EC funded project, in which I am participating with partners from Italy, Spain and Sweden, dedicated to comprehensive risk assessment in nuclear waste disposal based on geosphere kodelling, sensitivity analysis, model uncertainty quantification and parallel computing.

Environmental sampling

November 26th, 2-5pm RSS Headquarters, Errol St

Design in environmental research - a neglected topic?

John Jeffers, University of Kent at Canterbury

Abstract
We live and work at a time when data collection and statistical computations have become easy almost to the point of triviality.  Paradoxically, the design of data collection, never sufficiently emphasised in the teaching of Statistics and Computing, have been weakened by an apparent belief that extensive computation can make up for any deficiencies in the design of data collection.
Starting with an emphasis on the importance of defining the population about which we are seeking to make inferences, this paper reviews the requirements of sampling and experimental design in environmental research.  It stresses that sufficient attention to the design of environmental studies, with due regard to the difficulties of estimating population parameters and the interaction between critical factors, before any data are collected, is an essential pre-requisite for valid studies of environmental issues.

Spatial sampling strategies for  patch detection and surface prediction

Jon Barry, University of Lancaster

Abstract
Patch detection is important in several areas- most notably in pollution surveys of contaminated land, where the need is to find so called 'hotspots' of pollutants, and in shellfish surveys where the need is to find patches of shellfish above specified densities.  Of course, you want to be sure that if you don't find anything that you haven't missed anything worth finding!
Another important area is the prediction of surfaces based on discrete sampling of the surface and techniques such as kriging.  I suggest the use of 'blob' designs (ones with two densities of sampling) to try to provide good coverage of the area as well as good estimation of the covariance function at short distances.
 

Environmental Extremes
joint meeting with the North Eastern local group, Thursday 10th December at 2pm in Merz Court, University of Newcastle (travel information at  Newcastle ).

Speakers include:

Richard Smith, North Carolina.
Trends in rainfall extremes.

Jan Beirlant, Leuven University
Linear models in extreme value statistics.

Jonathon Tawn, Lancaster University
Modelling dependence for extremes of environmental time series.

Future meetings in 1999

Tues, March 16:  Statistics with GIS and Remote Sensing:  Some issues at the interface.
 

  10.00 to 17.00 at the Royal Statistical Society, 12 Errol Street, London EC1Y 8LX

This meeting aims to explore the uses and requirements for statistical approaches to spatially
referenced environmental data, particularly in the context of data typically stored in
Geographical Information Systems (GIS) or collected by remote sensing techniques. The speakers
include a mix of geographers and statistcians and discussion between all participants will be
encouraged.

10.00    Coffee

10.30    Useful statistics for the GIS User?
               Neil Stuart, Department of Geography, University of Edinburgh
11.20    Geostatistics for remote sensing and GIS
               Peter Atkinson, Department of Geography, University of Southampton
12.10    Registration of remotely-sensed images using an estimated digital elevation model
               Chris Glasbey, Biomathematics and Statistics Scotland, Edinburgh

13.00    Lunch

14.00    Generalized linear models and geostatistics: an application to mapping sea bird
densities on the North Sea.
               Edzer Pebesma, Department of Physical Geography, University of Utrecht
14.50    Some Developments in Multivariate Descriptive Methods for Spatial Data
              Trevor Bailey and Wojtek Krzanowski, School of Mathematical Sciences, University
              of Exeter

15.40     Tea

16.00    Filtering of multitemporal radar images for large scale mapping
              Shaun Quegan, School of Mathematics and Statistics, University of Sheffield

There will be a fee of £15, which includes a buffet lunch, coffee and tea.
 

SYNOPSES:

Useful statistics for the GIS User?

Neil Stuart (University of Edinburgh)

This talk will provide an overview of the areas of GIS analysis where statistical methods are
presently found and encourage the audience to assist in identifying useful methods that could be
applied by GIS users with real world problems to solve.    A series of examples on the topical
theme of environmental analysis with GIS will be used to identify some of the commoner
statistical methods presently applied in conjunction with GIS analysis.   These examples serve
to illustrate the kinds of statistics that have been operationalised to date and to reveal some
of the gaps between statistical theory and current practice.   Some observations are then
offered on the benefits that users report from applying statistics within their GIS analysis.
By combining these reported needs with a list proposed by Openshaw of GIS-able statistical
methods, the talk concludes by suggesting some areas where greater interaction with
statisticians could help the GIS user communities with their goal of producing information
products of a known reliability.

------------------

Geostatistics for remote sensing and GIS

Peter Atkinson (University of Southhampton)

Geostatistics has developed steadily over the last thirty years from essentially a technique for
spatial interpolation known as kriging to a sophisticated suite of statistical techniques for
the analysis of spatial data. Such techniques include indicator kriging and indicator
conditional simualtion, each of which will be discussed in the talk. Geostatistics was first
adopted by the remote sensing community over ten years ago. The focus of attention then was on
the identification of spatial structure in images and how to use parameterised models of the
(stationary) spatial covariance function to inform prediction of a variable of interest at the
ground given a sparse sample of it and the remote imagery. The logical solution was provided by
cokriging and the linear model of coregionalisation. However, cokriging has not been adopted
widely largely because of the difficulties posed by fitting the linear model of
coregionalisation. Cokriging will be discussed within a remote sensing context and some
alternatives will be reviewed.

------------------

Registration of remotely-sensed images using an estimated digital elevation model

Chris Glasbey (Biomathematics and Statistics Scotland)

Digital elevation models simplify the registration of airborne, remotely-sensed images.  If
unavailable, they can be estimated by minimising the sum of two terms: a measure of agreement
between the image and a map, and a thin-plate bending-energy.  A computationally-efficient,
multiresolution algorithm is proposed to solve the optimisation problem, with cross-validation
used to choose the relative weighting of the two terms.  The method is illustrated using a
synthetic aperture radar (SAR) image.

------------------

Generalized linear models and geostatistics: an application to mapping sea bird densities on the
North Sea.

Edzer Pebesma (University of Utrecht)

Geostatistical models can handle both non-stationary trends and spatially correlated residuals,
but they deal badly with count data or binary data. Generalised linear models on the other hand
are ill-suited for predicting spatial processes. Some recent papers have shown how the two
modelling approaches can be combined. An application of one such approach to the mapping of sea
bird densities on the Dutch part of the North Sea will be presented. I will also discuss how
this all links to GIS.

------------------

Some Developments in Multivariate Descriptive Methods for Spatial Data

Trevor Bailey and Wojtek Krzanowski (University of Exeter)

Currently there are few forms of statistical analysis specifically designed to handle spatial
structure and dependency both within and between spatially indexed multivariate responses, but
one such form is sometimes referred to as `spatial factor analysis' . This talk describes some
extensions to this group of techniques, which are primarily designed for use in exploratory
analysis and which are well-suited to the visual and interactive environment which GIS provides.

-------------------

Filtering of multitemporal radar images for large scale mapping

Shaun Quegan (University of Sheffield)

Effective use of radar images for large scale mapping requires methods to cope with a noiselike
effect known as speckle. This paper shows how multitemporal images can be combined in an optimal
fashion to reduce the speckle in each image in the temporal sequence, and how the requirements
of the particular problem may require further spatial filtering. The importance of first and
second order statistical properties on the choice of approach will be stressed and it will be
shown that different sensors require different approaches. The analysis and its application will
be illustrated in the context of mapping forest area and wetland rice.
 

April 8th 2-5pm,  Sheffield University,
Environmental Epidemiology.

A joint meeting with the Sheffield local group, to be held in the Hicks Building, LT2

Speakers are
Dr Lesley Rushton
Time series analysis - a limited methodology for quantifying individual risks from air pollution
 

The relationship between air pollution and adverse health outcomes is an important area of concern in the area of environmental health.
There are numerous published studies which have employed a time series approach to this issue and many more are planned or underway.  These
model the daily count of some health outcome, eg deaths or hospital admissions, with daily data on various air pollutants and climatic
variables, together with other potential confounding variables.  There is a general tendency for these studies to report a very weak positive
association between health outcomes and pollutants, but the relationships vary in their nature and strength. This presentation
will illustrate these points using the studies within the Air Pollution and Health: a European Approach (APHEA) project.  Although
the results from these studies imply evidence of causality they are limited in their ability to determine and evaluate the benefits of any
risk reduction strategies.  Study designs are needed which enable a more accurate estimate of individual exposures to be made and which
facilitate the investigation of the interrelationships both between the various air pollutants and between the pollutants and other
factors such as climate.
 

Dr Jon Wakefield
Bayesian modelling of disease risk around environmental point sources: methods and applications.
 

Recently there has been increased interest, from both  the media and the public, in the question: Is there  an excess of disease risk close to a pre-specified point source ?  To answer this question routinely-available public health data may be analysed.  In particular Diggle, Elliott, Morris and Shaddick (1997) have proposed  likelihood-based methods for dealing with such data using parametric forms to relate disease risk to distance from source.
 In this paper we propose a Bayesian approach to this problem. We extend the basic model of these authors to incorporate random effects.  These can be used for both the accommodation of extra-Poisson variation and, perhaps more importantly, for diagnostic purposes in order to suggest alterations to the basic location/risk function. We develop an informative prior distribution based on epidemiological considerations. The methods are illustrated using data from an investigation into the occurrence of stomach cancer close to five municipial solid waste incinerators.  The conclusions are compared with those obtained from a number of alternative methods.

 Diggle, P., Elliott, P., Morris, S., and Shaddick, G. (1997).
 Regression modelling of disease risk in relation to point sources.
 To appear in Journal of the Royal Statistical Society, Series A.

3.30pm Tea

Professor Alan Gelfand
Bayesian approaches for misaligned data layers and the modifiable areal unit problem.
  In attempting do establish relationships between spatial variables, one often encounters misaligned data layers. For example, the response variable may be observed on one areal grid and the explanatory variable on another. In some cases one variable is observed areally while the other is observed at point sources.  In attempting to reconcile such misalignment, one finds a variety of ad hoc methods in the literature.  We propose fully model-based approaches to develop such regressions.  Such models are necessarily hierarchical and provide full inference, avoiding typically inappropriate asymptotics associated with likelihood-based
approaches.  We discuss general approaches for handling the different types of misalignment and then illustrate the first situation (with which we have the most experience) using an  environmental risk dataset and also a very large deforestation dataset.
 

Study Group Meetings 1997/1998

September 23rd 2-5pm, Napier University, Edinburgh.

Biodiversity

November 19th, 2-5pm, Senate Chambers, Nottingham University

Discussion meeting on Spatio-temporal modelling

Speakers are

Dr P Blackwell, University of Sheffield

Multivariate Diffusion in a Random Environment

A multivariate diffusion process in a random temporal environment can be used as a flexible model of movement, e.g. in interpreting behavioural and radio-tracking observations on wildlife. Monte Carlo approaches to bayesian estimation for such a process will be considered, with the analysis of data on wood-mouse movement.

Prof P Diggle, University of Lancaster

Space-time calibration of radar-rainfall data.

The talk describes an analysis of data consisting of time-series of rainfall measurements from am irregular spatial network of sites in the North-West of England, together with time series of radar reflectance values from a fine pixel grid which covers the region spanned by the rainfall sites. The objective is to estimate the rainfall intensity on the fine grid.

The approach taken is:

to develop state-space time-series models for the relationship between radar reflectance and rainfall intensity of a single site;

to investigate the empirical spatial correlation structure of the state variables as estimated by these single-site models;

to integrate the single-site models into a spatio-temporal state-space model for the entire network.

Prof K V Mardia, University of Leeds

The kriged Kalman filter

A model which combines effectively two well established approaches of spatial statistics (kriging) and time series (Kalman filter) will be described for spatio-temporal modelling. Practical applications of this model will be highlighted.

Dr M Mugglestone, IACR-Rothamsted

Spectral analysis of spatio-temporal data

Spectral analysis provides a convenient framework for spatio-temporal modelling. The methodological feasibility of the appraoch will be contratsed with difficulties in obtaining suitable data fro conjoint spatial and temporal modelling in ecological and agricultural studies.

December 9th, RSS headquarter, Errol St, London

Joint meeting with GAS, at 4.30pm

R McNally, LRF Centre for Clinical Epidemiology, Leeds

A review of disease mapping methods with application to leukaemia incidence data

Abstract

As part of a recent study of the descriptive epidemiology of leukaemia and related conditions in the UK (Cartwright, McNally et al, 1997), the geographical distributions of these diseases have been mapped. Different mapping approaches are discussed in relation to these data. In particular, both empirical and fully Bayesian methods are used to obtain relative risk estimates. The mapping of these estimates is illustrated using the leukaemia data.

and

N Best, Imperial College School of Medicine, London

Ecological modelling of health and exposure data measured at disparate spatial scales.

Abstract

This paper considers an ecological (area-level) regression analysis of the effect of socio-economic deprivation and road-traffic pollution on respiratory disorders in children. Bayesian hierarchical models with Markov random field priors will be used to address statistical problems such as sparse event data and spatial autocorrelation. An extension to these models will then be described which attempts to accommodate individual-level and fine-resolution geographical covariates by relating all observable quantities to an underlying continuous random field model.

January 14th, Errol St, London at 5pm

RSS Ordinary Discussion Meeting

Alternatives to Economic Statistics as Indicators of national well-being

Statistical issues in developing indicators of sustainable development

John Custance and Hilary Hillier, Dept of Environment, Transport and the Regions

Sustainability indicators-integrating quality of life and environmental protection

Roger Levett, CAG Consultants

February 26, Errol St, London at 2pm

Joint meeting with Biometrics Society on Sustainability

Speakers include Roger Payne and Kevin Stokes

The assessment of agricultural sustainability

Roger Payne, IACR-Rothamsted

The sustainability of farming systems is recognised to be a very important issue, but there is less agreement about how this should be measured and assessed. The talk will describe experiences and conclusions from a project commissioned by the Rockefeller Foundation to study sustainability using data from long-term experiments at Rothamsted and five other agricultural institutes world-wide. It was shown that economic considerations could successfully be measured by indexes such as Total factor productivity. Externalities such as ecosystem health, however, were found to be more difficult to assess.

The estimation and utility of biological reference points as used in precautionary fisheries management

Kevin Stokes, CEFAS

This talk will provide background on marine fisheries management with emphasis on certain recent attempts to ensure sustainability by application of a precautionary approach. Marine fisheries systems are complex and uncertain. Defining appropriate criteria for overfishing is not straightforward and estimating adopted criteria and their error levels is fraught with difficulty. The talk will highlight the main classes of criteria that are used and will discuss their properties and utilities.

14th May 1998, RSS headquarters, Errol St

A joint meeting with the NOVARTIS Foundation

Environmental Statistics: Analysing data for Environmental Policy.

Programme:

10.00 Registration and coffee

10.30 Welcome by Gregory Bock of the Novartis Foundation

10.35 Chairman's introduction by Vic Barnett of University of Nottingham

10.45 Operational evaluation of air quality models by Paul Sampson, University of Washington

This talk addresses the analysis and modelling of spatiotemporal field observations of tropospheric ozone for the purpose of the assessment ot operational evaluation of complex photochemical air quality models. We begin with a discussion of issues in the evaluation of gridded model predictions of pollutant concentrations against point field measurements, recommending an approach that is essentially inverse to current procedures recommended by the US EPA. We briefly review a number of approaches to modelling the complex, temporally and spatially non-stationary covariance structure of the field measurements and use such a model as a basis for computing optimal grid-cell estimates. We discuss the comparison of such estimates with predictions of the 'SARMAP' air quality model for the San Joaquin Valley in California. We also discuss a proposal to compare the spatial covariance characteristics of model predictions with the spatial covariance structure of field monitoring observations.

11.25 Coffee

11.45 Statistical examination of environmental regulatory standards for ozone by Larry Cox, US EPA

Recent attention in Europe on setting environmental regulatory standards, most visibly a study by the UK Royal Commission on Environmental Protection, has emphasised the value of basing standards and related compliance criteri on statistical principles. These concers are documented in a recent book by Barnett and O'Hagan (1997). They are timely for consideration in the USA where periodic review of air quality standards has led to revision of US national regulatory standards for ambient ozone and particulate matter. Salient statistical issues include accounting for errors for the first and second kind due to sampling and measurement error. Even though these issues can entail potentially significant practical consequences, and are routine statistically, they are often absent or not prominent in regulations. This paper examines these issues through a statistical analysis of past, current and proposed standards for ambient ozone, based on intensive monitoring in California's San Joaquin Valley during summer 1990 performed under the SARMAP project.

12.25 Measuring and modelling pollution exposure for risk analysis by Jim Zidek, University of british Columbia.

The greta scale and complexity of environmental risk analysis offers major methodological challenges to those engaged in policy-making. This paper will describe some of the problems that arise in measurement and modelling and sloutions that have been proposed.

The first problem to be addressed will be that of imputing unmeasured pollutant concentrations. This special kind of measurement error pronlem arises because of the high cost assocaited with installing and maintaining monitoring stations. We describe statistical solutions even where several pollutants are of concernm not all stations monitor all pollutants, the stations were established at different times and finally the stations aggregate their concentration measures over different time scales.

Even with these imputed measurements in habd, the problem of inferring actual exposure remains. For example, in very hot weather people will tend to stay inside and population levels of exposure to say ozone could well be below thsoe predicted by the ambient measurements. Setting air quality criteria should ideally recognise the discrepancies likely to arise. Computer models that predict actual exposures from ambient levels will be described.

The final major problem to be addressed is that of inferring the often extraordinarily subtle impcats of these exposures. The case of ozone and its relation to respriatory morbidit exemplifies the problem. We describe solutions to the problem of combining impact assessment over space and time to assess that association. We indicate how these solutions can incoprorate those for the problems described above. Some recent findings that raise questions about air quality criteria in use will be presented.

13.05 Lunch

14.00 Statistics and environmental policy: standards Australia and the Victorian EPA by Geoff Lasslett, CSIRO

Environmental objectives are statements of policy which need to be written in such a way that they are capable of being assessed using information from a monitoring programme. An environmental monitoring programme has to be adequate in its quality and quantitiy of data so that the environmental objectives can be assessed. Also, the resulting data should be able to contribute information towards decisions to modify the policy at a later time if necessary.

A key intermediate role exists for the use of statistical inference in providing a logical framework for using monitoring data to test hypotheses about fulfilment of environmental objectives. The undertaking of a statistical inferential approach influences both the development of testable environmental objectives and the design of monitoring programs. Some case studies from the Victorian EPA monitoring programs will be presented to illustrate these points.

14.40 The use and non-use of statistics in setting up the UK's policies for protecting us from water pollutation by Tony Warn, UK Environment Agency

It is a fact, largely unappreciated, that hardly any statistical know-how has been built into much of the legislation by which Europeans protect their water environment. With few exceptions, the rules for using data to measure the need to act are inflexible, and they can be inept. The outcome can be confusion and worry about the state of the environment, and the potential for investment that is wasted and untimely.

Brushing aside the debate on why statisticians are not invited to contribute to issues so obviously within their territory, this paper covers the problems with the legislation abd how we work round them.

Many of the Directives issued by the European Union fail to distinguish between the concepts of populationa nd sample. The main issues are

that failure is defined in terms of volatile face-value estimates of summary statistics;

instances where quality must be shown to improve or not to deteriorate; and

standards that are expressed as absolute limits.

We conclude that assessments of compliance and performance should be done in duplicate. the first appraisal must follow the stipulations of the Directive; the second should be a proper statistical estimation. The results from the proper assessment should be sued to colour how we report those in the first. We should point out when failure was statistically significant and when the Directive's rules precluded a good evaluation of the real need to take action.

The paper goes on to draw lessons from recent national policy on Consents for Dsicharges, River Quality Objectives and Directives, where the aim is to take proper account of statistical fundamentals.

15.20 Tea

15.40 Integrating data for sustainable development by Jeanette Heycox, Australian Bureau of Statistics

In an information rich world, organising frameworks that focus data to reveal succint views and inter-relationships are essential to allow society to grapple with the complexity of the world around us. The goal of sustainable development requires a particularly large range of data covering many disciplines-economic, environmental, social and institutional. In the last decade, numerous frameworks have been developed for organising data to inform debate and analyse management options for sustainable development. Sustainable development is nebulous in nature, the field of study is borad, the number of inter-relationships immense and the undersatnding of what constitutes sustainable development is in its infancy. It is not surprising then that frameworks developed to date have had limited success. However, each one has advanced our understanding by providing a new perspective and fresh insights. It is with this aim that I present a new framework for sustainable development.

The distribution of resources (DOR) framework explicitly quantifies the relationship of changes in the quantity and quaility of our natural and produced capital to the policy concerns of: supply side efficiency, demand side efficiency, intra-generational distribution of quality of life attributes and population size. The framework is moderate in its data requirements, hence it should be applicable to a broad range of nations with varying statistical capacity. It reveals that sustainable development has a multiple range of solutions with numerous trade-offs.

16.20 Chairman's closing remarks

Registration (including lunch, refreshments and documentation) is 10 pounds (payable to Royal Statistical Society).

Study Group Meetings 1996/1997

Venue

Contact: Marian Scott, Dept of Statistics, University of Glasgow, phone 0141 330 5125, e-mail: marian@stats.gla.ac.uk
  Future meetings April 17 1997

Handling Uncertainty and variation in the setting of Environmental Pollution Standards.

Vic Barnett and Tony O'Hagan, University of Nottingham

Venue 12 Errol St at 2.00pm

Abstract

Environmental pollution standards effectively ignore statistical variation and uncertainty, being expressed as statements about population parameters (ideal standards) or about sampling procedures (realisable) standards. The former often give no guidance on compliance, the how (or what) to infer from the outcomes. A hybrid approach: the statistically verifiable ideal standard is needed. Such a case was recently made to the Royal Commission on Environmental Pollution
  May 20 1997, Lancaster University

Adaptive estimation of trends and seasonality in environmental time series.

Peter Young, Lancaster University

Abstract

The paper will describe recent advances in the recursive extrapolation, interpolation and smoothing of nonstationary time series, with particular application to environmental systems. The basic approach is formulated in temrs of an Unobserved Component (UC) model with stochastically defined trend and periodic components, although it is straightforward to add other stochastic components, such as though arising from the effects of damped cyclicity and other exogenous variables. This UC model is particularly useful for adaptive seasonal adjustment, signal extraction, interpolation over gaps, and forecasting ( or backcasting). The Kalman Filter (KF) and Fixed Interval Smoothing (FIS) algorithms are exploited for estimating the various components, with the noise variance ratio and any other hyper-parameters associated with the stochastic state-space models of the components estimated by the optimisation of a least squares cost function defined in terms of the difference between the logarithmic pseudo-spectra of the UC model and the AIC identified autoregressive model of the time series. In all of the applications considered, this novel cost function not only seems to yield improved convergence characteristics when compared with the alternative ML cost function, but it also has reduced numerical requirements. A number of practical environmental examples will be discussed, including the analysis of very long temperature series in the USA and Antarctica; and the modelling of non-linear rainfall-flow processes.

Modelling global temperatures

John Haslett, Trinity College, Dublin

Abstract

Under the auspices of the UN's International Panel on Climate Change, researchers at the Climatic Research Unit at the University of East Anglia have assembled a data set on temperatures since the mid 1850's. These data provide much of the smoking gun evidence that the Earth is warming and that it may continue to warm. They also provide a detailed test bed to examine many detailed hypotheses, particularly those associated with the global circulation models which attempt to model the fluid dynamics of the Earth's atmosphere. This paper will introduce the data set and discuss some of the empirical modelling attempted by the author.

The data are pseudo-data and correspond to a complete coverage of the Earth in 5 degree boxes for each month since 1854. It is clear that the data collection effort 140 years ago was very patchy, with almost no coverage of the oceans or the southern hemisphere; indeed not until very recently have there been satisfactory coverage of the oceans. However, very considerable effort has been expended by the IPCC in cleaning the data, adjusting for changes in data capture methods and in extrapolating the available data to the entire surface of the planet. This talk will discuss the data.

The analysis presented will concern the empirical modelling of these data. The broad features of the variation will be discussed; for example both the covariance and the magnitude of the trend exhibit strong seasonal structures.

This research is 'work in progress'.

Venue Room B58/B59, George Fox Building, University of Lancaster

Time: 2pm This page is maintained by Marian Scott