Make your work easier and more efficient installing the rrojasdatabank  toolbar ( you can customize it ) in your browser. 
Counter visits from more than 160  countries and 1400 universities (details)

The political economy of development
This academic site promotes excellence in teaching and researching economics and development, and the advancing of describing, understanding, explaining and theorizing.
About us- Castellano- Français - Dedication
Home- Themes- Reports- Statistics/Search- Lecture notes/News- People's Century- Puro Chile- Mapuche


World indicators on the environmentWorld Energy Statistics - Time SeriesEconomic inequality
--------------------------------------------------------------------------
RRojas Databank Journal/ January 1997
--------------------------------------------------------------------------
SCIENCE, RISK ANALYSIS AND ENVIRONMENTAL POLICY DECISIONS
by John M. Stonehouse and John D. Mumford
UNITED NATIONS ENVIRONMENT PROGRAMME
The United Nations Environment Programme was launched by the UN Conference on the Human Environment, held in Stockholm in 1972. Its mandate is to catalyze and coordinate activities to increase scientific understanding of environmental change and develop environmental management tools. Among its milestones over the past two decades is the creation of Earthwatch to gather, analyse and convey information about the state of the global environment. In the form of environmental management tools, UNEP's efforts have led to conventions to protect stratospheric ozone, to control the transboundary movement of hazardous wastes and to protect the planet's biological diversity, among others.
ENVIRONMENT AND TRADE SERIES
UNEP's Environment and Trade Series is intended to provide both trade and environmental decision-makers with background analysis of various policy issues of relevance to the trade and environment debate. Views expressed in this paper are not necessarily those of the United Nations Environment Programme. To obtain additional free copies of this article, please write to:
Environment and Trade
UNEP
15, chemin des Anemones
CH-1219 Chatelaine
Geneva, Switzerland
Fax: 41.22.796.9240
Series Coordinator: Scott Vaughan
FOREWORD

The 1992 "Earth Summit" found common ground upon which human development can be put on an environmentally sustainable footing. In 1993, completion of negotiations for the Uruguay Round set the course for a further liberalisation of international trade. One of the most pressing and complex challenges facing our generation is the search for a workable synthesis of the two, of economic relations and environmental realities.
We must embark upon this course, not because it is easy, but because it is necessary. Our planet's ecological vital-signs continue to warn us of an accelerating rate of degradation -- depletion of the ozone layer that shields us from harmful solar radiation, erosion of productive soils needed to grow food, contamination of freshwater with hazardous wastes, depletion of fish stocks, the massive loss of biodiversity, the threat of climate change and global warming.
An important challenge identified at the Earth Summit is ensuring that trade and environment are "mutually supportive." It is hoped that this series, providing analysis on selected environmental issues of relevance to the environment - trade debate, will contribute to the search for solutions now underway.
Elizabeth Dowdeswell
Executive Director

THE AUTHORS

This paper was prepared for the United Nations Environment Programme by JOHN M. STONEHOUSE and JOHN D. MUMFORD, of the Environmental Law Group of the Imperial College Centre for Environmental Technology.
John Stonehouse studied history at Oxford University and zoology at Imperial College, London University. He was awarded a PhD degree at the Imperial College Centre for Environmental Technology, for a study of the economic and social context of pesticide use by smallholder farmers in Colombia. As a researcher and administrator in the science and policy of the environment and development, he has also worked for the European Commission in Brussels, and in the commercial and nongovernmental sectors.
John Mumford is a Senior Lecturer in the Department of Biology and the Centre for Environmental Technology at Imperial College, London. He studied at Purdue University, USA and Imperial College, London, and has worked extensively in developing countries. He teaches courses in resource management, pest management and decision analysis. His research interests cover economics, risk assessment and public policy in pest management and related fields of resource management.
 
ACKNOWLEDGEMENTS

The authors wish to thank Dr Helen ApSimon, Martin Hession and Steve Hollins, of Imperial College, and Professor Peter Calow, of the University of Sheffield, for useful information and comments. All opinions expressed, and responsibility for any errors, remain with the authors.
 
The Environmental Law Group (ELG) of the Imperial College Centre for Environmental Technology (ICCET)
The The Enviornmental Law Group was formed as s distinct research and teaching group within Imperial College of Science and Technology and other University of London colleges, and is dedicated to advanced legal research in the areas of national, European and international environmental law. The Group's location within Impmerial College gives it a distinct ability to work with leading scientists and engineers, and input from other disciplines features very highly in much of its activity.
For further infromation contact:
Environmental Law Group Centre for Environmental Technology Imperial College of Science, Technology and Medicine
48 Prince's Gardens
London SW7 SPE
United Kingdom
Phone: 44 71 589 5111 (Ext.: 7220)
Fax: 44 71 823 7892

CONTENTS
 
1. Introduction
- Problems in the Interpretation of Scientific Data for the Making of Environmental Policy
- Issues of the Philosophy of Science
- Testing Hypotheses in the Probabilistic Sciences
- The Environmental Sciences
- Conclusion - Categorization of Environmental Risk
2. Risk Analysis
- Risk Assessment
A. The Hypothetical Case of "Fully Probabilized Risk"
i. Hazard Identification
ii. Assignment of Probabilities
iii. Consequence Modelling
B. Breakdowns of the Fully Probabilized Ideal, and Tools Used to Manage Them
C. The Use of Expert Judgment
- Risk Evaluation
- Conclusions
3. Environmental Risk Analysis
- The Toxicology of a Single Species
A. Hazard Identification
B. Dose: Response Assessment
C. Exposure Assessment
D. Risk Characterization
E. Risk Management
F. Post Facto Monitoring
- Ecosystem Toxicology
A. The Assessment of Ecosystem as Represented by Individual Species
B The Holistic Assessment of Ecosystems
- Nontoxicological ans Systematic Stress
- Conclusions
- Paradigms for Ecological Risk Assessment
4. Conclusion
5. References
--------------------------------------------------------------------------
1. INTRODUCTION

PROBLEMS IN THE INTERPRETATION OF SCIENTIFIC DATA FOR THE MAKING OF ENVIRONMENTAL POLICY

The worlds of science and policy are not well suited for mutual communication (Stewart, 1993). In the sorts of information which they provide and require, the attitudes and priorities of their practitioners, and in their operational vocabularies, including the meaning given to the same words, the fields of science, economics, law and politics differ in subtle and sometimes frustrating ways. This paper addresses the uncertainties and imperfections of scientific information for the shaping of environmental policy, and how a suite of decision tools, under the broad title of "Environmental Risk Assessment", is being developed to assis decision making in the light of these imperfections.
Environmental risk assessment tools, therefore, have arisen as a response to the shortcomings of scientific information for policy purposes, and so it is in these shortcomings themselves that the logical roots of the field lie. These shortcomings extend throughout the field of environmental science, from the fundamental philosophy of science, through the inferential nature of the probabilistic sciences, to the multidisciplinary complexity of the environmental sciences.
 
ISSUES OF THE PHILOSOPHY OF SCIENCE

The philosophical problem at the heart of science is known as Hume's Problem: stated briefly, this is that no matter how many times a phenomenon is observed, we cannot be sure that this represents a universal pattern or "law" - logically, even large numbers of past examples cannot guarantee that the same events and relationships will continue to be found in future.
Even when "laws" provide convincing and effective explanations for phenomena, we cannot be sure that such "laws" will continue to hold true, nor that we have not imperfectly devised them and that new exceptions will not arise in circumstances which differ in ways we have not foreseen. The most successful approach (so far) to Hume's Problem, that of Karl Popper, does not "solve" it, but allows scientific theories to be tested. This approach takes advantage of the logical asymmetry that, although a theory cannot be proved, it can be disproved, by the finding of contrary examples.
No matter how many times we observe the sun rise in the East, we cannot prove the statement "the sun always rises in the East" to be true; yet one observation of it rising in the West (or anywhere else) will disprove it.
In Popper's approach, therefore, a good scientific theory must be phrased so that it may be shown to be wrong, or falsified, by the finding of examples which contradict it, and so, although not verifiable, it is testable, by systematic attempts to disprove it. Under this view falsifiability is the hallmark of scientific statements, distinguishing them from non-scientific ones, such as those arising from tradition, emotion or authority. This allows competing theories to be compared one to another, and useful ones (i.e. those that appear to hold true) are employed with a provisional and qualified approval, until such time as they may be disproved or, more often, refined and improved as exceptions to them are found, investigated and incorporated. This qualified and provisional nature of the acceptance of scientific theories, which therefore can never be proved as "laws" in the popular sense of universal, unbreakable and true generalizations, is not always understood by the layman who may demand certainty of the scientist, and may be uncomfortable with information which, instead, although part of a field of theories which are continually improving by a rigourous process of comparison and testing, is only "the best generally applicable thing we have found so far". In fact, however, this causes few problems in the real world: scientific "laws" are more commonly evaluated, for practical purposes, by their usefulness than by their truth. The confidence we may have in the usefulness of "laws" builds up, over time, as they are seen to be of service in practical applications.
A second philosophical problem at the heart of science is that two objects or phenomena, such as two pieces of steel loosely describable as "centimetre cubes" cannot be shown to be identical - even with the most refined means of chemical, physical and electromagnetic measurement, we can never be sure that the two do not differ in some subtle way, either because they differ on a scale so fine it is beyond our powers of measurement, or because they differ in some unfamiliar way of which we are ignorant.
However, just as a theory can never be proved but can be disproved, they may be shown to be different: measurement may reveal differences in size, composition or properties.
These two asymmetries - that theories cannot be proved, but can be tested by attempts to disprove them if they are falsifiable, and that two things cannot be proved to be the same but can be proved to be different - lie at the heart of the approach taken to the testing of scientific hypotheses. A hypothesis is "a supposition put forward in explanation of observed facts" (Uvarov & Isaacs, 1986), and their testing is the core of science. For example, one may wish to determine whether the trout in a polluted river are thinner (or fatter) than those in an unpolluted river nearby. The fish can be weighed or measured. The hypothesis "One trout population is thinner than the other" cannot be falsified, as they cannot be shown to be the same; the alternative, however, "The two trout populations are identical in weight" can be falsified (by showing them to differ) and can therefore be tested, and so it is this latter hypothesis which the scientist will test. If one were sure that the only possible effect of interest were that the fish in the polluted river would be thinner, then a similarly testable hypothesis would be "The trout in the polluted river are the same weight or fatter" - a so-called "one-tailed test". However as, in fact, any anomaly is likely to be operationally relevant, two-tailed tests, whereby the tested hypothesis is "They are identical in weight", allowing deviation to be shown either way, are much more widely used.
 
TESTING HYPOTHESES IN THE PROBABILISTIC SCIENCES

All sciences are in a sense probabilistic, that the statements which they may produce are not certainties but "best guesses", but some are more probabilistic than others. In particular, the biological sciences have to address the problem that organisms differ, in their genes and in the effects wrought on them by their environments. For example, the trout populations of any two rivers will be different, due to inherent variability. This creates serious problems for the testing of hypotheses such as the one outlined above - if the two populations differ, how can we conclude that those in the polluted river are thinner or fatter than those in the unpolluted one due to a meaningful difference between the rivers and not due to the random effects of inherent variability? In addition, in a case such as this the fish populations will be sampled, measurements being made of a selection of individuals, rather than the entire population, and the extent to which the sample genuinely represents its parent population will vary. These thorny problems are approached by the large and powerful set of tools known as statistical analysis.
The essential function of statistics in this sense is to assess variation due to different causes - in our case the inherent variation in the two trout populations and that between the populations in the two rivers - and to formulate the probability that the variation source of interest (that between the rivers) is so large, relative to the general variation, that it may be considered to be "significant". In our example, our hypothesis that "the two populations are identical" may be tested by a statistical tool (such as the versatile, simple and ubiquitous t-test) which will produce a conclusion such as "there is a 97% probability, taking account of the general variation in both populations, that this hypothesis is wrong" - i.e. we can say with 97% confidence that the two populations are not the same.
The usefulness and relevance of such conclusions are limited in several ways, which in turn are addressed by a suite of techniques.
First, the power of statistical tools to provide significant results depends, more or less, on two things. First is the size of the sample taken - one can see intuitively, as well as mathematically, that the more samples are taken, the greater will be the confidence in the conclusions made. As a result, any test will fail to produce a significant difference if its sample size is too small; on the other hand, any two populations will, in the real world, differ due to inherent biological variability and so, conversely, a statistically significant difference may almost always be obtained if the sample size is large enough, regardless of whether or not the difference has any ecological meaning in the real world.
Second, statistical tools differ in their power to resolve differences. As a rule, the more powerful tests rely on certain properties of the data (principally how the tested characteristics are distributed among the populations) meeting certain conditions, which must be assumed to be the case. (These conditions are expressed as "parameters", and tests of this type are called "parametric statistics"). For the test to be reliable, these conditions must hold, and so the employment of the test assumes that they do. Tests can be carried out to compare the observed conditions with some of these assumptions but, as these tests employ hypotheses which themselves are falsifiable, they can, like the central hypothesis, indicate that the conditions are not the same as those assumed, but not that they are the same: the correctness of the assumptions can never be proved, allowing the validity of a parametric test to be questioned. When these assumptions are not considered valid, alternative test techniques, with less demanding assumptions ("nonparametric statistics"), may be used, but such tests have lesser resolving power than their more demanding parametric cousins: for example, the nonparametric Mann-Whitney test has about 95% of the power-efficiency of its parametric analogue, the t-test (Siegel & Castellan, 1988).
This has real-world implications: the same data may be tested by two different tests, and a significant effect (at any level) found with a more powerful and demanding test, but not with one which is less so - the meaningful interpretation of the importance of the test results will therefore require a knowledge of which test was used, why, and how its power compares with that of alternatives.
Second, the principle of cause is not necessarily demonstrated. The detection of a significant difference between the two populations cannot automatically be attributed to pollution - the rivers and their populations may differ in other ways, even in ways undetected. In fact, the common-sense notion of "cause" is, like all theories, impossible to prove to exist in any case, because of Hume's problem, and the inference of cause is another large field of scientific philosophical study. Statistical techniques can provide evidence of the strength of the association between two variables, but such associations can mislead: a famous example is that urban fox populations in Britain are closely associated with the political representation of areas, being much more abundant in Conservative areas than in Labour ones - these two phenomena are clearly not causing each other, by some bizarre mechanism such as foxes somehow bringing out the Conservative vote, but both are to be explained by a third, less obvious variable, composed of social and economic factors. (For many years, supporters of the tobacco industry argued that the association between smoking and lung cancer was such a "proxy association").
Tools exist to help the inference of causes. One of the most powerful of these is experimentation. This has three central advantages over nonexperimental observations. First, the causes whose effects are under investigation can be precisely applied by the experimenter. Second, their effects can be evaluated with reference to "control" cells where conditions are, as far as possible, identical to those in the test cells, except solely for the influence of the causal candidate under study, advancing considerably the ability to infer causation. Third, experimental and control cells can be repeated or replicated, increasing confidence that observed effects are not random. Convincing experimental results rely heavily on the care with which conditions other than the test or "forcing" variable are controlled and maintained, and the training of scientists emphasizes this as a crucial part of "good laboratory practice". In nonexperimental research, logical tools, such as a systematic consideration of alternative explanations, can be employed (Blalock, 1961). However, despite the usefulness of such tools, the inference of causation, like all scientific "laws", of which it is in effect a particularly useful and important subset, causation can not be conclusively proven.
Essentially, therefore, the convincing demonstration of hypothesized causes depend on two things. The first is plausibility, or common sense. The second is, as for all "laws", the usefulness of the relationship in practical, real-world applications, in which confidence builds with use over time.
This is important because understanding of likely causes is central to the meaning and usefulness of science. When causal mechanisms are understood with some confidence, scientific information can be used with greater precision and flexibility, and the importance of the inference of causes is a recurring theme in the pages below.
Resolving these two issues, the use of statistics and the inference of cause, depends heavily on the skill and competence of scientists. In the case of a failure of elementary principles of scientific good practice, its detection by the scientists' audience is difficult. Two additional techniques are used by the scientific community to address these problems. The first is the concept of replicability - results are presented to the scientific audience with sufficient detail of the experimental and analytical methods to allow another researcher to repeat the experiment or observation exactly, and thus to confirm its findings. This principle has a somewhat archaic air, and tends to the publication of scientific papers filled with long and, to the casual reader, boring descriptions of materials, equipment and procedures; but its central point is a critical check on scientific competence, allowing contested reports to be checked by any interested party.
The second is the concept of peer review, whereby articles submitted for publication are anonymously checked by established practitioners in the same field, allowing their judgement to be used to check that of the author in important decisions such as the formulation of hypotheses, choice of statistical tests and interpretation of results.
Third among the practical limitations of probabilistic scientific procedures is that the asymmetry of the formulation of the hypothesis, as described above, does not permit the contrary hypothesis to be tested. A failure to reject the hypothesis that the populations are the same is not to conclude that they are the same. Scientists often treat "not significantly different" as equivalent to "the same", but this is inconvenient, to say the least, for many policy applications - politicians and the public may wish to know how sure we are that the two trout populations are the same, but a confidence value for this cannot be given.
Fourth, the significance level itself is subjectively chosen. When presenting findings to their peers, by means of scientific journals, scientists traditionally accept the level of 95% to assign "significance".
Some argue that this is too low. If, for example, twenty hypotheses are tested, and all rejected by the criterion of just exceeding 95% confidence, then, on average, one of these findings (1 in 20) will simply be wrong. On the other hand, for policy purposes 95% may be considered too high - many people would not be happy if a 92% probability of the existence of an adverse effect was dismissed as "not significant". This is, at root, due to the different objectives of scientists and policy makers. The scientist seeks the accurate generalization of findings. The policy maker seeks a level of confidence reliable and robust enough to achieve social goals in a variety of cases, and the high political cost of a single error produces a tendency to caution. The social and political evaluation of a significance level is a subjective area where statistical probabilities found by science must be evaluated in a social context.
These latter two issues, which limit the usefulness of scientific results to real-world decision makers, have led to efforts to define "significance" other than the 95% criterion of the rejection of the hypothesis of sameness. Approaches are advocated to a search for "environmental significance". One is that differences, instead of merely being assessed for significance, be quantified wherever possible. This may be done by calculating expressions of statistical confidence in the sizes of the differences between populations; or by "regression", a statistical technique which fits a mathematical description to the relationship between two or more variables, for example between the weights of fish and the concentrations of pollutants in their habitats. Another approach is the replacement of the tested hypothesis "these two populations are the same in weight" with that "the difference in weight between these two populations is delta or larger", where delta is a preset value of environmental significance and meaning, and the outcome sought is that, by the rejection of this hypothesis, the difference is shown to be less than delta. Adoption of such techniques, ideally simultaneously in several countries in a harmonized way, may go far to enhance the facility with which nonscientist decision makers can digest statistical scientific information.
The above is a brief summary of the scientific and statistical tools used for the inference of conclusions in the probabilistic sciences. Two issues arise.
First, the output of the scientific method is complex, not always obvious, probabilistic and heavily dependent on the details of the methods employed, and the skill of the scientists in choosing and executing them. For these reasons it is important that the workings, assumptions and limitations of these tools be known when scientific information is being presented to decision makers and policy formers. To take an extreme example, an unscrupulous scientist, working for a company suspected of causing pollution, may wish to fail to show that fish in two polluted and unpolluted rivers were significantly different, in order to support the case of his employer: in such a case, failure to find a significant difference could be rigged, by the use of a weak test or of a small sample size. Understanding of statistics would be needed to detect this.
Second, this scientific output is not always presented, or indeed presentable, in a way which to the layman obviously has economic, social or environmental "meaning". The interpretation of the policy significance of scientific findings is particularly complex in the environmental sciences.
 
THE ENVIRONMENTAL SCIENCES

Turning to the environmental sciences, the complex, probabilistic and provisional nature of statistical data is amplified. For the making of policy, a critically important issue is the difficult and questionable concept of "ecological health" - how is the health of an ecosystem to be defined or assessed, in either scientific or ethical terms? These problems may be considered by a progression through the levels of ecosystem organization.
The smallest point of ecological measurement is generally the individual organism. The health of an organism may be measured in many ways - by its death or survival, or by its behaviour, reproductive success or susceptibility to disease. These multiple points, which in the case of human beings are well known from the field of medicine, require different and imaginative methods of analysis. In a sense, individual organisms are therefore not the smallest point of ecological measurement, which can be carried back to organs, such as the liver, brain or blood supply, or to less tangible aspects such as contentedness, irritability or dissatisfaction. When the value to be protected is the individual, as is most evident in the case of human beings, but also in intelligent, "charismatic" animals such as elephants or dolphins, these values need to be assessed.
Above the organism in the scale of ecological organisation is the population - the group of organisms, of the same species, which routinely interact and interbreed. Populations, and the species which together they compose, are often the value whose protection is desired. The health of a population is to be measured most obviously by the survival of its members.
The impact of damage on population survival is experimentally assessed by "dose:response" techniques. These use regression methods to fit mathematical descriptions to the percentage of the population killed or affected at points along a scale of increasing levels of stress, typically doses of poison (or "toxin"). Dose:response analysis is an important field of "bioassays" - the use of living material for the measurement of effects.
From the derived relationship one can calculate the doses likely to affect any fraction of the population, and the toxicities of chemicals are traditionally compared by the use of the dose which kills 50% - the 50% Lethal Dose, or "LD50". This is because, as in all statistical techniques, the confidence of the estimates is greatest for those values which are central in the range of values tested, which is the LD50 if all doses are tested between those which kill no test organisms and those which kill them all. The confidence is much lesser at the extremes of the distribution, which is awkward because it is generally the lower extreme, the dose which kills very few individuals or none at all, which is of importance for policy. For this reason, and concern at the distress caused to large number of animals in conventional LD50 testing, attention has now shifted to tests at the lower limits of toxicity, particularly to the identification of the largest dose which causes No Observed Adverse Effect - the NOAE Level, or "NOAEL". Toxicity tests clearly can hardly be carried out on humans, or on many other animals, either because they are endangered (the platypus), large and unruly to work with (the bison), charismatic animals whose suffering is painful to contemplate (the chimpanzee) or, more usually, all three (the white rhinoceros, the blue whale). In such cases, estimates must be extrapolated from similar or related species, which is not always easy (as for whales). Extrapolation is the inference of values outside the scale which has actually been measured, by the extension of relationships inferred where it has been measured. It is a recurring, critical and contentious issue in environmental risk analysis, and will be treated further.
Beyond relatively simple estimates of individual survival other criteria come into play. A population may provide an ecological function of value to society, such as fish or game which provide food, trees which protect a watershed from erosion or insect predators which kill farm pests, and in such cases these functions may be more important than the population per se. For conservation, the long term viability of a species is related to its genetic diversity and this, assessed by techniques to measure the diversity within the population of the structure of proteins or DNA, is a widespread conservation goal, which has policy implications such as the attachment of greater importance to the maintenance of a small, isolated population of a species than to a similar but larger one, as the total genetic diversity is more threatened by the reduction of the former.
Beyond the protection of populations is that of their physical environment. As well as by environmental toxins, this can be disturbed by, for example, temperature changes (as in the water above and below dams) and soil erosion. Some of these effects can be subtle, and they vary between locations and circumstances - for example, eroded soil, whose loss from the land is itself deterioration, becomes a pollutant in rivers, blocking sunlight and affecting concentrations of dissolved gases such as oxygen.
Environmental complexity is dramatically enhanced at the level of ecosystems. The disruption of ecosystems is hard to predict, taking place across an array of trophic levels (i.e. stages up food chains, from plants to herbivores to predators) and ecological interactions. First, tests cannot be carried out on all species present, or even on an extrapolable analogue of each. Second, pollutants themselves can and do move up food chains, often by a process of "ecoaccumulation", by which toxins accumulate in the bodies of predators, eaten in their prey, which can exceed concentrations in the prey themselves, leading to unforeseen problems, such as the effects of ecoaccumulated DDT in bird predators. Third, influences on one species can have effects on others through their ecological interactions, by the process of "ecological mediation", often "knockon effects" reverberating up and down food chains - predator and herbivore populations will suffer from losses of their food organisms and, conversely, populations may expand following losses of their competitors (as Southern Ocean seals and penguins following the virtual elimination of whales) or of their predators (as Californian sea urchins following the near-extinction of the sea otters which eat them, which in turn led to decimation by the sea urchins of kelp beds). The assessment of ecological effects, even post facto, relies on the careful selection and sampling of species, and of indices of species diversity and function, such as energy budgets - the choice and measurement of these variables requires considerable scientific expertise as well as clear political priorities as to the "ecological value" of concern.
Finally, studies of the global environment rely on the addition of atmospheric and oceanic sciences. Although some such phenomena can be experimentally assessed, as by the release of harmless substances to evaluate the likely behaviour and dispersal of plumes of toxic gases or liquids, their evaluation on a global scale depends on mathematical models which are currently variable and imprecise. For example, estimates of the impact of industrial carbon dioxide on the Greenhouse Effect vary widely, causing proposed levels of reduction to lack consensual scientific validity (Rotmans & Swart, 1990), and allowing the leaders of industrialized and petroleum-producing countries to question the existence of the Effect.
The concept of "ecosystem health" implies an analogy with the health of an organism, such as a person. In fact, this analogy is imperfect and may be misleading. "Health" implies a stable, equilibrium state, to which a system tends to return, by self-regulating processes developed by evolution, as when an animal such as a person recovers from disease or injury. This process of control helps us to define "health", but there is no evidence that ecosystems do regulate themselves in this way - that there is an objective "optimal state" which may be used for defining ecosystem health (Calow, 1992).
The original, "normal" state of an ecosystem, before any human disturbance, may be used to define its "health", but ecosystems are themselves dynamic, in constant flux rather than a steady state. This makes the concept of "normality" itself hard to assess, as changes in ecosystem structure and function occur naturally, and whether and which changes in ecosystems may be attributable to human disturbance are hard to demonstrate with conviction (Barlow et al., 1992). One approach to this problem is, where possible, to discover the limits of historical fluctuations, and to aim not to exceed them - this approach has been used to define "acceptable" fluctuations in wildlife populations around oilfields (Maki, 1992) and in the temperature of the planet. Yet many, if not most, ecosystems are in fact the product of inevitable human disturbance. This raises two issues. First, the original state may not be knowable - archaeological research has shown that even the Amazon basin, apparently a "wilderness", has been shaped by human communities for centuries. Second, many human-dominated systems also have ecological functions - agricultural fields, for example, are highly "unnatural", yet have values as ecosystems which we would wish to protect.
In the event, the properties which we use to define the qualities of ecosystems which we wish to protect are almost always chosen by human value judgements - such as biodiversity or natural beauty. In general, people tend to be conservative, wishing merely to keep ecosystems as they have come to know them, i.e. as they are now. For example, the rough grazing common lands of southern England, since the abandonment of their traditional use for animal grazing, are being invaded by small trees such as birches, which used to be held in check by the animals. Wishing to maintain the heath ecosystems, themselves rich in wildlife, "conservation" volunteers spend weekends uprooting these trees, and yet their arrival is merely the first step in the return of these areas to their natural state before human disturbance - mature deciduous forest - which is therefore prevented by society because of the higher value that it places on familiar than on original ecosystems in this case.
It is also important to distinguish between change and harm in the disturbance of ecosystems. Harm is, as discussed above, largely defined by subjective human value judgements, and not all changes may be harmful in this sense. For example, if a population of beneficial insects such as honeybees is exposed to a pesticide but, after initial reduction of the population, a resistant type emerges, so that the population recovers to its initial levels and remains unaffected by the pesticide thereafter, ecological change has clearly taken place but many would argue that no harm has been done. On a larger scale, the global extermination of the smallpox virus received little condemnation from conservationists, and the extermination of other "pests" such as, say, the Anopheles mosquitoes which transmit malaria, might generally be seen by society as a good thing.
 
CONCLUSION - CATEGORIZATION OF ENVIRONMENTAL RISK

This bewildering complexity and interaction of environmental components has led to the concept of "Environmental Risk Analysis" being employed in different ways, and "Environment" itself to mean different things in science and in policy making. As is the nature of environmental phenomena, these interact and overlap, but a categorization can be discerned:
1) Toxic threats, whether chemical or by radiation, to human health through the environment, often called "environmental risk" but now, because of confusion, increasingly being referred to as "environmental health risk".
2) Toxic threats to the natural environment, or ecosystems. Threats to humans through the environment and those to ecosystems, the latter now coming to be called "ecosystem risk", have suffered in the past from being too infrequently distinguished, though this is now increasingly emphasized (Norton et al., 1992). The distinction can, however, be overstated (Barlow et al., 1992), as the two have many points in common, as human beings are at the top of many food chains, and so environmental toxins attack ecosystems and people together, as in the Minamata Bay tragedy.
3) Nontoxic threats to ecosystems, such as construction projects, for example of roads or dams, land clearance for agriculture or the removal of individuals from a population by hunting, fishing or treecutting.
Toxicology, the study of the action of toxins, comprises the bulk of environmental risk analysis, as toxins are perhaps the best known and most disconcerting of the stresses man applies to the environment, and some authors (e.g. Calabrese & Baldwin, 1993) treat ecosystem risk as almost entirely toxicological. The extent to which these and nontoxicological equivalents can be treated in consistent ways is debatable. Generally speaking the two may be distinguished in legal terms, for setting of standards. Chemical environmental threats, being found wherever the products are used or transported, are not precisely controlled in their distribution, and thus generally addressed as products, whose use and concentrations are controlled on the market by standards. Nontoxic threats are generally site-specific, being effectively caused by processes, and thus limited to their sites, such as nuclear power stations or logged forests, and are regulated by local monitoring and evaluated by the site-specific subdiscipline of environmental risk assessment known as Environmental Impact Assessment. Nonetheless, this distinction is not perfect: a chemical manufacturing plant will pose site-specific toxic risks, such as possible leaks of intermediary chemicals used in the manufacturing process, quite different from those encountered in the use of its products, and the risks from the introduction to an environment of an alien species will be found wherever the species may spread. Also, with regard to trade implications, it can no longer be argued that product standards, affecting goods traded, are the sole concern of national environmental policy as process standards affect only the host country: increasingly states are reluctant to import products, however satisfactory their intrinsic qualities, that are produced with serious environmental process risks (Charnowitz, 1993).
There is therefore a case to be made for treating toxic and nontoxic environmental dangers as essentially similar. In the USA, the Environmental Protection Agency (USEPA) is developing framework guidelines for such a harmonized approach, arguing, sometimes implicitly, that ecological "stressors" can be treated alike, with "extents" of nontoxic factors, such as the number of inches drop in the water level of a wetland, or the percentage of trees in a forest felled, treatable as essentially like doses of chemicals (Norton et al., 1992). The distinction between product and process standards remains a useful one, however, and can be maintained in discussions over the international harmonization of environmental risk analysis processes.
A more meaningful division of the field may be into the evaluation of those environmental threats which are essentially incremental in their effects, such as most toxins, or the erosion of the ozone layer, whose effects may be considered in some way additive, and those which are systematic, a single event producing widespread repercussions, such as the extermination of a species, the introduction of an exotic species, or the construction of a dam. Incremental stressors may be essentially transient, the effects of each encounter with them declining over time, as in the case of a pesticide which decomposes relatively quickly, in the environment or in the bodies of organisms, or cumulative, as in the case of a pesticide which decomposes only slowly, and therefore may build up, over a series of encounters, in organisms or their environment. Especially if they are transient, they can to some extent be controlled after their action, as monitoring may reveal when their build-up approaches a critical level; the risks of systematic stresses must be considered a priori, and often without the option of laboratory or small-scale testing. Ecosystems may be able to recover from incremental stresses, such as by "negative feedback" processes whereby an increase in a phenomenon itself produces ecological responses to lower its levels back towards an equilibrium; systematic stresses may trigger "positive feedback" whereby the departure from equilibrium accelerates itself, so that relatively small initial departures can spiral uncontrollably into a collapse of the system.
A critical question for the risk analyst is often whether a specific phenomenon may be incremental or systematic, or may switch from one to the other. For example, selective forestry in North America may be incremental in its effects, as natural tree regeneration restores the forest to equilibrium; the same approach in Tropical America may be systematic as, even if few trees are removed, the opening of the forest by access roads often permits the settlement of the land by colonizers, who clear the remaining trees for agriculture, bringing about a collapse of the system. More seriously, the contribution of anthropogenic carbon dioxide to global warming has apparently been governed by negative feedback processes so far, as not all the carbon dioxide released since the industrial revolution is still in the atmosphere (though nobody knows quite where it is); but in future, under some scenarios, such as the shrinking of the boreal forest belt as it retreats towards the North Pole or the large scale melting of ice caps, these processes may change to those of positive feedback, the effects of global warming producing further warming, with serious consequences for the global environment. Partly because of the large scientific uncertainties surrounding these processes, the consideration of whether environmental threats are likely to be incremental or systematic and which, if any, feedback processes may apply, are critical questions for the management of environmental risk. "For risk to a phenomenon as complex as the environment, it can never be certain that the process of hazard identification has been sufficiently imaginative and rigourous."
 
2. RISK ANALYSIS
 
"Risk" is defined many ways, in different technical fields and among laypeople. In general, it is the probability of something bad happening. The something bad is known as "damage" - defined formally as "the loss of inherent quality suffered by an entity" (Warner, 1992) or, more generally, as "something you would like to avoid".
"Probability" itself is a somewhat slippery concept. The probability of an event happening is the mathematically expressed chance of it doing so, usually as a fraction of 1. To be meaningful, probabilities have to be bounded, usually within a certain period of time, such as before the end of the week, or in association with a particular event, such as a single toss of a coin: unbounded probabilities tend to be 1, or certainty - it is certain that all life on planet Earth will end one day, but the probability of it happening this year is smaller. They are also generally conditional, the probability of events, such as risks, varying with circumstances, such as between populations exposed or susceptible to them to different degrees.
For example, the probability of being hit by a car while crossing the road in Los Angeles, California, is higher than that in a small rural village.
As formally expressed in mathematical terms, probabilities are useful in the real world because they can be manipulated by arithmetic - added, subtracted, multiplied and so forth.
Probabilities are generally approximated by frequencies: if a coin tossed 1000 times comes up "heads" 500 of them, we estimate a probability of 0.5 that it will do so next time. However, probabilities cannot be truly known, for the same reason as that behind Hume's problem - we have no way of knowing that the next 1000 tosses will not all be "tails". Nonetheless, over a large number of replications frequencies tend to represent probabilities. The frequencies of rare events, on the other hand, are little use for the estimation of probabilities. For example there has never been a recorded earthquake in Manchester, England - but there remains a small probability that there will be (Stewart, 1990). In a sense, therefore, all probability estimates are subjective.
"Hazard" is an inherent capacity to cause damage - an intrinsic property. DDT therefore simply is hazardous, but this is a property, and so is not probabilistic and therefore not risky - risk is applied to events, not properties. Risk is the probability of the potential of hazard becoming realised as damage.
Risk can be defined in at least ten ways (Pidgeon et al., 1992), but in risk analysis two definitions are most common - risk is either "the probability of an adverse event" or "the probability of an adverse event multiplied by the extent of the damage caused." The former is the formal economic definition, and widely used, but the latter is gaining currency.
The former, under which the latter is called "detriment" (Warner, 1992), will be followed here. The reason for this is that the quantified assessment of detriment implies the quantification of damage, and in much environmental risk assessment this quantification, implicit though it may be, is not performed.
The quantification of environmental values is a field in itself.
Damage, and therefore risk, can broadly be categorized as economic, safety or environmental - in order of decreasing ease of quantifiability. Economic risk is a well worked field, having being studied for centuries by the profession of its analysis, the insurance industry. Much progress has been made in the quantification and even economic valuation of safety risks too, by estimation of the prices people are apparently willing to pay for specified reductions in the risk of injury or death. Economic valuation of the environment is performed and advocated, on the grounds that "any decision implies valuation" and that if environmental valuation is not tried in the taking of an environmentally relevant decision "we do not know whether the decision was a sound one or not in terms of economic efficiency" (Barde & Pearce, 1991). Environmental assets can be valued in essentially two ways. One is variants of "hedonic pricing" using real world data, by estimating values from, for example, house and hotel prices which may be expected to reflect preparedness to pay for proximity to environmental resources, or from spending on travel to visit them. The other is variants of "contingent valuation", which estimate how much people are willing to pay for environmental assets by asking hypothetical questions (Barde & Pearce, 1991). However, the valuation of the environment is not necessary for environmental risk assessment and is, in fact, only rarely carried out. The ways in which environmental values are considered are various and subtle, and will be discussed below.
In contrast, the benefits of taking risks are commonly more easily valued - the financial benefits from logging a forest, using a pesticide or building a nuclear power station are usually worked out by the people who propose to do it. It should be remembered that to avoid risk has its costs - pesticides, timber and energy are valuable and useful things. The decision of whether or not to take risk is one of balance.
All action, including inaction, carries some risk, and so all decisions entail risk judgement, even if subconsciously (Calow, 1994; Stewart, 1993). "Risk analysis" is a more explicit, systematic and communicable extension of this. It risk comprises two interconnecting parts. "Risk assessment" is the structured assembly of an estimate or representation of the total risk from an activity. In principle, it is scientific, deterministic and objective. Its results are then used for "Risk evaluation", the consideration of the significance and meaning of risk, used for making the decision of whether the risk is to be taken. Risk evaluation is necessarily subjective, political and value-laden. Assessment and evaluation together form the whole process of "Risk analysis".
 
RISK ASSESSMENT
 
A. THE HYPOTHETICAL CASE OF "Fully Probabilized Risk" The concept of risk as the probability of an event is most readily conceived in the contemplation of accidents, the province of Engineering Risk assessment (Crossland et al., 1992). A typical such assessment looks at the sequence of events which may lead to a system failure in a factory, leading to a leak of toxic chemicals to the environment. This process, the assembly of a composite representation of risk or "risk model", entails the sequential consideration of the events which, happening one after the other, together finally manifest themselves as the total risk.
 
i. Hazard Identification
The process begins with "Hazard Identification", the formal listing of what hazards may exist and what risky situations may arise, and how. For the analysis of an engineering plant, this entails a systematic examination of plant components, functions and interconnections. For environmental risks, involving more complex and loosely bounded systems, the process is more a question of imagination, to consider problems which may arise, within an attempt to bound the exercise to prevent the consideration of risks spreading ever-outward to the limits of the environmental field.
 
ii. Assignment of Probabilities
Hazards which are identified as both serious and realistic are broken down into the sequential events which may bring them into effect, for example the release of chemicals from a system may require the failure of a valve, followed by the failure of a pump, followed by the failure of a control system and of a warning alarm (or the failure of a human operator to recognise the alarm and to act on it). The "Assignment of Probabilities" to each of these events is performed in appropriate ways - for example, the failure rate of a certain valve or pump may be well known from records of its performance in other systems (large technical databases exist of this information), others may be estimated from expert opinions. The probability of the final, disastrous failure, is derived by the multiplication of the component probabilities. For example, if every ten years failure "A" has a probability of 1 in 20 (i.e. 0.05) and failure "B" one of 1 in 100 (i.e. 0.01) the probability of their both occurring at once is 0.05 x 0.01 = 0.0005 (i.e. 1 in 2000).
 
iii. Consequence Modelling
From the derivation of the probability of an accident, the analysis proceeds to "Consequence Modelling". At this point the assessment begins to lose analytical rigour. The dispersal of pollutants may be affected by the wind or water currents at the time, to be derived from meteorological information or models of flows in estuaries or the atmosphere; the population of organisms exposed will be site-specific and, in cases such as humans at work or home or migratory animals, may vary in time, and this must be recorded. These techniques are used to estimate the likely exposure of a population. Following on from them, the uptake of pollutants by organisms must be estimated, both biophysically, whether through the skin, gut or lungs, and behaviourally if, for example, birds are likely to eat contaminated insects, worms or fish. From the projected uptake values the likely toxic damage must be estimated by dose:response information, from LD50 or NOAEL values, these relationships generally estimated by extrapolation from tests on laboratory organisms such as rats.
Environmental risk assessment is not always assessing the likelihood of an accident: the use of agricultural pesticides, for example, is risky but deliberate. In this case, the assessment is of risk not in the sense of an accident happening, but of something having effects that cannot be precisely foreseen, and consequence modelling comprises the whole operation. The tools used for their analysis are similar because we do not know what will happen - risk analysis is therefore in a sense "ignorance analysis".
This analysis can be assembled in several ways to provide a holistic picture. At the simplest level, the probabilities of certain levels of damage can be calculated as single number values, by the multiplicative combination of probability estimates at each step.
For the decision maker to make a more informed judgement, a more detailed picture of the chain of events may be needed. One way to do this is by means of branching "trees" of sequential failures and their ramifications. "Event trees" follow the possible consequences onward from an initial failure; "fault trees" proceed the other way, moving backward up the chain of causes needed to result in a specified undesirable outcome, the dramatically named "top event" (Crossland et al., 1992).
A more sophisticated synthesis may be in the form of a computer model. Such models express, in a quantified way, the flow of events or material through the system. They are particularly useful, and widely used, for consequence modelling, using information about the flows of air or water, and the passage of food up food chains through ecosystems, to estimate likely movements of toxins through the environment and ecological compartments. However, they require careful validation and testing, and confidence in their reliability is generally built up only over time, as their relevance to the real world is shown (Barlow et al., 1992; Geraghty, 1993).
From such a synthesis, risk management has options for action. At the extremes, it may be decided that no risk is present worth considering further; or that the risk is unacceptable, and the construction of the plant, sale of the pesticide or whatever, simply forbidden (in fact, in either of these cases, such a decision may be reached before completion of the full risk assessment process). Otherwise, the process allows risk managers to identify critical points on the trees where preventative measures may be focused to reduce risk, such as the addition of another safety system or the doubling up of control personnel where human alertness and competence is of importance.
 
B. BREAKDOWNS OF THE FULLY PROBABILIZED IDEAL, AND TOOLS USED TO MANAGE THEM
In reality, allocations of probabilities to top events can only very rarely be assessed with even the level of accuracy implied above. These are only some of the areas of uncertainty.
1) For risk to a phenomenon as complex as the environment, it can never be certain that the process of hazard identification has been sufficiently imaginative and rigourous. This is important because the hazards identified dictate the whole of the subsequent structuring of the risk model, and so the model may be inappropriate if hazards are unforeseen, arising out of, say, ecological interactions or unexpected conditions such as violent weather. Tools exist to facilitate this. For example a well established routine for chemical plants, "HAZOP", which involves a dissection of the process pathway into all its discrete steps, and the application to each step of code phrases such as "more than", "less than" or "other than" to force consideration of how the process may deviate from intention, has been modified, as "GENHAZ", to structure a similarly systematic exploration of the potential hazards from the experimental release of genetically modified organisms to the environment (Royal Commission, 1992). Even with such tools, however, the need remains for imagination and for judgement - in particular, the exercise must be logically limited to prevent the endless expansion of "the universe of discourse" (Crossland et al., 1992).
2) In sciences such as biology and meteorology, where the data to be worked with are both variable and influenced by a large number of factors, the very measurement of variables is often fiendishly difficult. For example, bioassays performed in supposedly identical ways by different laboratories may produce widely different results without strict control of a plethora of conditions and variables. This can be investigated and controlled by "ring tests", whereby the same test is performed by a group of laboratories with identical test materials, and their results compared.
The usefulness of this, and the dangers it may illustrate, was shown by a ring test carried out in 1986 by the European Community (EC) of toxicity to laboratory clone populations of Daphnia magna (a small freshwater shrimp widely used as a test organism for laboratory bioassays; a clone population is one in which all individuals are genetically identical, being derived asexually from a single parent). Among the 22 laboratories which reported results, the conclusions varied by the maximum amount possible (i.e. the estimated toxic doses ranged from below the minimum concentration specified by the test procedure to above the maximum). Subsequent investigation found that not only did the clones of D. magna used differ between laboratories (including one case where a clone deviated, apparently by spontaneous random mutation, to produce a new genetic type) but also so did their rearing diets and conditions. Research found that both clones and rearing conditions did produce differences in the toxic susceptibility of populations, often in subtle ways - for example, susceptibility differed in shrimps whose mothers were fed different diets (Baird et al., 1989). It is for the discovery of problems like this that ring tests are carried out, and efforts have been and are being made, in the EC and elsewhere, to standardize procedures to avoid them, but they remain a disturbing source of potential inaccuracy.
3) The accuracy of statistical estimation decreases away from the centre of a probability distribution (usually the average or "mean") towards its edges, or "tails". This is inconvenient because it is at the tails that human interest is concentrated - particularly in the probabilities of very unlikely but very serious incidents, or disasters.
This is because in human perception the probabilities of risk are clearly not treated arithmetically: people will not, for example, treat a certain loss of $1, a 1 in 10 chance of a loss of $10 and a 1 in 1000 chance of a loss of $1000 as equally undesirable - if they did, nobody would ever insure anything. The estimation of the probabilities of rare but severe events is even less reliable than that of others.
4) In the assembly of probabilistic risk models probabilities are assumed to be independent from each other, yet may not be so: in the case of accidents, if a human operator fails to adjust a piece of equipment correctly, the same operator may fail to notice an emergency warning, due to tiredness, negligence or alcohol consumption; in environmental systems risks may, as outlined above, become systematic if positive feedback processes are activated.
5) Another assumption too often made is that the response of the natural world to stresses is linear, that is, that the response to an incremental increase in stress at one level will be related (albeit perhaps in a logarithmic, multiplicative way rather than an additive one) to the response at other levels. Yet this is not always the case: above threshold levels the system may respond in quite different ways. This applies in toxicology (for example, oxygen is highly toxic at high concentrations) and also in ecology, where work with mathematical models has shown that even simple ecological systems may, when disturbed beyond narrow boundaries, show violently erratic, chaotic behaviour. Increasingly, however, scientists are moving away from the assumption of linearity, although this greatly increases the complexity of analysis.
6) Yet another assumption often made is that the effects of processes are independent, so that when more than one is encountered together their effects are additive. This assumption, however, the base of the approach that complex processes may be reduced to their components, or reductionism, is often mistaken, interactive effects being commonly found. As with the assumption of linearity, scientists are moving beyond this assumption, however, although, again, this increases analytic complexity.
7) The processes of extrapolation are ubiquitous in environmental risk assessment, and lie at the heart of many of its difficulties. For ecological risk assessment, toxicological extrapolation is needed within species, to estimate the toxic susceptibility of the population from the sample tested; between species, from the species tested to those likely to be exposed; from laboratory to field conditions; from acute to chronic effects; and from effects in tests to those likely in real ecosystems. For example, extrapolation is needed to assess the likely effects of pulsed or episodic as opposed to steady-state exposure (Seager & Maltby, 1989); of cumulative build-up of effects (Spalding & Smit, 1993); of bioaccumulation of toxins in the tissues of organisms, or ecoaccumulation such as that of DDT in the progression up food chains; and the effects of simultaneous exposure to multiple stresses, whether these will be additive or interact in mutually reinforcing ways.
Because of their ubiquity and difficulty, much research effort has been spent on extrapolation techniques. In particular, attention has been focused on the search for understanding of the mechanisms in toxic processes - of how toxins are absorbed, distributed, modified and take effect. However, this search for understanding of the causes of toxic effects has produced little success - there are currently few "laws" inferred about toxic processes which are not merely extrapolations derived from observations of correlations (Barlow et al., 1992). Studies of parallel tests, to investigate the reliability of extrapolation between species, have been carried out in many cases, but have produced conflicting results. For example, between species reliable extrapolability has been found between some species with some chemicals, but not others (Emans et al., 1993; Greig-Smith, 1992). Similarly, progress is being made in the extrapolation from laboratory to field conditions (Munawar et al., 1992) and from acute effects to chronic (Giesy et al., 1989). Interactions, where two stresses encountered together exert greater stress than the sum of either alone, have been found with pesticide combinations in some species but not in others (Johnston et al., 1994). A good deal of work is needed for understanding the extrapolability of these effects and, indeed, it seems reasonable to suppose it will never be reliable.
8) This is just one example of the general situation that mechanisms are poorly understood. The problems of the convincing demonstration of causality have been discussed above, and are particularly severe in the complex and variable world of the environment. Some tools can assist the inference of causation even from nonexperimental observations, such as "falsification routines", the use of structured approaches (not unlike GENHAZ) to evaluate alternative possible causes for effects. For reliable prediction of environmental effects in any but the most general ways, for example for the building of precise computer models of systems, understanding of mechanisms is required, but is not well advanced (Calow, 1993a).
9) The complexity of environmental variables and their interactions through effects such as ecological mediation are another source of uncertainty in the interpretation of environmental risk information. Data may be of different types, from bioassays, computer models, published scientific papers or engineering or meteorological data, whose contributions must be balanced and evaluated. Inconsistencies in data sets, such as higher toxic susceptibilities in some species than others or accumulation in certain organs more than others, must be interpreted. The scale and scope of such information make its interpretation as much of an art as a science.
The outcome of all the problems raised above is twofold.
First, the results obtained often cannot be quantified. Instead, they may be qualitative only, or categorized - for example into "high", "medium" and "low" risks. Categorized conclusions lend themselves to understanding, recognise the inherent uncertainty of the knowledge and allow cross-comparisons and standardization, but they are themselves imprecise by their nature, and difficult to define in a standardizable way (GreigSmith, 1992). Findings of this sort have their uses, but their significance must be assessed with skill.
Second, the assessment of environmental risk is the evaluation not only of risk, which is a situation whose outcome is unknown but where the probabilities of each possible outcome are known, but also of uncertainty, where the probabilities themselves are unknown. Uncertainty is principally addressed by the use of "safety factors", whereby levels estimated to be "safe" are multiplied to provide those that will be treated as "safe", to reflect lack of confidence in extrapolations and so forth (Talcott, 1992).
In some fields the practice, and the factors used, are well established, for example the traditional but widespread factor of 100 applied when extrapolating rat toxicity data to humans. Yet even at this level other factors are not widely accepted (e.g. the proposed analogous factors of 500 for reproductive toxins and 1000 for carcinogens) (Barlow et al., 1992), and factors for other areas, such as ecosystem risks, are even less so. An alternative approach is uncertainty analysis.
Uncertainty analysis is by its nature an imprecise field, and indeed the expression "uncertainty analysis" may be considered something of an oxymoron. However there are tools which can be employed in situations of uncertainty. Chief of these is the attempt to express the likely degree of error of an estimate. Often these can only be verbal, but attempts have been made to draw up scales of such verbal descriptions, with rough analogues of quantified error estimates, for example "extremely uncertain" to indicate "probability of 0.025 that the true value is more than 100 times the estimate or less than 1/100 of it" (Talcott, 1992). The attachment of such estimates, rough as they may be, to each probability in a risk model helps to identify where the uncertainty arises, so that further research or particular efforts at risk reduction may be applied to them. For practical risk management purposes, another useful tool is sensitivity analysis, whereby highly uncertain values may be hypothetically considered to be in truth higher or lower than estimated, and the impact of these different values on the whole risk assessment considered. When the risk model is fully quantified and in the form of a computerized mathematical model, this may be done by the random insertion, one after the other, of large numbers of varying values for uncertain estimates, and the distribution of the final estimates evaluated, in a process poetically known as "Monte Carlo analysis" (Talcott, 1992). Nonetheless, uncertainty is sometimes irreducible, on the principle that "there are things that we know we don't know, but others that we don't know we don't know".
 
C. THE USE OF EXPERT JUDGEMENT
In the light of these sources of error and uncertainty, very often the only solution possible is the application of expert judgement. The employment of judgement, which is often essentially subjective, the "hunch" of an experienced practitioner, may seem something of a retreat from the rigours of scientific discipline, but in practical terms is not only inevitable, as many risks cannot be formally quantified, but productive - the experience and judgement of experts is, by its nature, generally not quantified, formalized or published, but remains valuable and relevant information, and to fail to make use of such expertise would be to fail to use all the relevant information available.
Given this situation, attempts are being made to formalize and rationalize the use of expert judgement. Its integration with more formal scientific methods is uneasy and uncertain, as certain principles of scientific research (such as replicability) are clearly undermined - indeed, it was largely to remove the need for such subjective judgement that many elements of the scientific method, such as replicability and the formal statement of falsifiable hypotheses, were developed in the first place. The application of some discipline and rigour to its use is important, both for the credibility of conclusions and also for the free transferability of assessments and their conclusions for the harmonization of processes in the interests of open trade.
The central tool for the rationalization of expert judgement is documentation. All authors in the fuzzier areas of environmental risk assessment are in agreement over the importance of a careful, explicit and comprehensive "document trail", in which the decisions reached and criteria used are clearly described and justified, so that the process can be evaluated, per se or for its subsequent applicability to other cases.
The employment of subjective probability values can be made coherent and useful by the employment of Bayesian probabilities. This approach is a way of combining probability information from different sources, incorporating, for example, old data or those which are considered to be relevant but not strictly applicable (such as fieldwork from different but similar geographical areas) so that the use of relevant information is maximized and none is lost or wasted. Additionally, new information as it arrives can be incorporated to upgrade the composite probability values, and the information incorporated can be weighted (for example by giving lesser weights to uncertain sources or those whose relevance is questionable), to give appropriate emphasis to those component information sources which are more or less reliable or important (Dillon & Officer, 1971).
The problems of the maximization of the objectivity of experts can also be addressed, at the simplest level by the widespread use of contractually independent consultants rather than inhouse teams, or by more sophisticated approaches such as "Delphi techniques". Delphi techniques allow the views of a group of experts to be canvassed and synthesized, by the polling of the individuals, usually by a questionnaire, the circulation to them of a synthesis of the results, and then repeating the process, allowing comment and debate, until a consensus is reached or, at worst, disagreements have been systematically explored and the reasons for them made clear and explicit, all usually carried out under anonymity for each member, to allow free expression and criticism.
The technique has advantages over a physical gathering of experts in that all views are heard, neither the majority nor a vocal minority dominating the debate, and that it is generally cheaper, as simultaneous attention by all parties is not needed and travel costs are saved. Sampling research has suggested that as few as ten members can comprise an effective group, as long as they are carefully chosen to be representative. The role of subjective input is not eliminated, however, as the effectiveness of the approach depends on the skill and judgement of the central "monitor team" who choose the participants and questions, and write the syntheses (Richey et al., 1985a).
The demand for expert judgement for environmental risk assessment is large and growing and, although educational systems are expanding to meet this, is likely to exceed supply for the near future. This problem, together with the desire to standardize and harmonize expert findings, has led to interest in the use of expert systems, computer models of human expertise which, relative to a real person, are cheaper and faster and, being deterministic information processors, produce consistent and therefore standardized conclusions. The snag with these systems is that their construction, requiring the formal expression of subjective and intuitive inference processes and exhaustive subsequent checking of the output, is very slow, taking several years. However, several have been in development for some time, and their arrival in operation may substantially increase their use (Geraghty, 1993).
 
RISK EVALUATION
There comes a point where the assessment of risk hands over to the process of its evaluation: at this point we reach the limits of science and the need for social and political values. The need for subjective expert judgement in the face of the limits of environmental information, described above, can be termed "assessment subjectivity" - in theory, it is subjectivity of estimation, and not value-laden; it is therefore different in principle from "evaluation subjectivity", which expressly weighs ethical and social values.
The cleanliness of the distinction between assessment and evaluation is not absolute, the subject of much debate, and currently in decline. This is for two reasons. First, progress in the psychological and social sciences is beginning to shed some light onto the meaning and derivation of "subjective" risk perception. Second, the essentially subjective nature of the use of expert judgement in the process of risk assessment, while it can be reduced by methods such as Delphi techniques and expert systems, cannot be eliminated. The significance of the latter point has been shown by anthropological and sociological studies of communities of scientists themselves, which have shown their assumptions, opinions and beliefs to be shaped by the institutions and conditions in which they develop.
Scientists can be shown to tend to share bodies of knowledge, accepted procedures, attitudes and confidence in institutions among themselves more than they do with other people, which has caused environmentalists often to view the use of their opinions with suspicion. The public reaction to discoveries about the safety of nuclear facilities in the former Warsaw Pact countries has shown how quickly the validity of the opinions of scientific bodies can be discredited when political circumstances change (Pidgeon et al., 1992).
However, the assessment-evaluation distinction, at least as a concept, for the understanding of the risk analysis process, is valid and useful. It is broadly analogous to that between the functions of technocratic "civil servants" and representative "politicians": in principle, at least, the evaluator of risk should be politically accountable, whereas the assessor is "simply doing a job".
The evaluation of risk is a direct function of its perception. Risk perception has been studied for some time by behavioural scientists. Its assessment is a bit like environmental valuation - a choice between researching "expressed preferences" from peoples' answers to hypothetical questions and "revealed preferences" from actual behaviour, such as willingness to work in risky jobs for higher wages, implicitly valuing risk, with the latter currently losing credibility to the former.
The central conclusion is that risk perception is multidimensional, perceptions differing between individuals and contexts, so that for policy purposes "risk" as rated by the public is not reducible to a single value such as a function, albeit a subjective one, of probability x damage (Pidgeon et al., 1992). As a result, risk policy making inevitably involves confrontation, balance and compromise. People also differ in their willingness to take risks, apparently as part of inherent personality diversity.
Risk may not be "accepted" and therefore "acceptable" in the sense that it is consciously balanced against a perceived good, but "tolerated" where necessary: in general people are risk averse, which has led to caution in the acceptance of risk in policy making. This, together with the imprecision of the safety factors by which it is commonly expressed, has led to frustration among many scientists that overcautious risk avoidance is leading to losses of social benefits through worthwhile risks not being taken, and some predictions that "de minimis" risk analyses are losing credibility and will not be sustainable (Berry, 1990). However, the view that the public "fails" to appreciate true risk levels and that public intolerance of "small" risks, that scientists consider to be worth taking, may be addressed by education and explanation of real risk levels in rational terms, is currently losing ground to the view that risk perceptions are deeply seated. Therefore, public confrontations may be better averted if these perceptions are not treated as "problems" to be "rectified" by education but as, in themselves, valid elements of the background to the public management of risk, which should be addressed and incorporated from the beginning in risk analyses. The "acceptability" of risk is therefore widely estimated using entirely political processes, being subjectively set by politically accountable bodies such as parliaments. Despite its imprecision, this in fact works adequately well as the best available way to allow public opinions of acceptability to be used (Pidgeon et al., 1992).
A well known finding about risk perception is that several factors, other than the absolute probabilistic risk level, influence peoples' perceptions of risk as intolerable. These "outrage factors" have been extensively researched for safety risks, and reasons for some of them investigated. In general, risks are less tolerable if they are:
- with inadequate, unclear or selective corresponding benefits;
- imposed, not being undertaken voluntarily by the risk bearer;
- outside personal control, the risk bearer having to trust to others the management of risk;
- seen as unethical or unfair in the distribution of the risk burden;
- publicly managed by untrustworthy information sources;
- artificial as opposed to natural;
- insidious, with damage happening in unseen ways (e.g. poisoning);
- of unknown time duration, particularly slow-acting damage which may affect subsequent generations;
- unfamiliar;
- associated with memorable events such as disasters. (People appear to have a peculiar dread of mass-scale disasters; also it appears that an event is conceived as "probable" if it is easily imagined or remembered, due to a convenient internal "rule of thumb" to facilitate the building of mental models of the world, known as the "availability heuristic" - memorable events such as disasters are by definition mentally "available" in this way [Pidgeon et al., 1992].)
In fact, most of these criteria seem intuitively apparent to the layman. The importance of the trustworthiness of institutions, for example, is readily appreciable, and it has been argued that failing public trust in state organizations is an endemic threat to public consensus on risk analysis, arising from inherent properties in current political and media structures, amplified by rapid social and technical change (Slovic, 1993).
Clearly, some risks fulfill these criteria more than others.
Recreational mountaineering, for example, is an extremely dangerous activity, yet widely popular because it entails so few outrage factors - for example, the benefits in enjoyment value are considerable and risks are under personal control, voluntarily undertaken and not insidious (as they basically entail falling off mountains). By contrast, environmental risks, whether health risks such as air pollution or food additives or ecological risks such pesticide use or the risks of oil tanker accidents, rate highly on many of these criteria: environmental risks are particularly prone to outrage factors.
Beyond the psychological study of individual risk perception stretches the relatively new research field of how it is influenced by social and cultural structures. Some of this work, such as suggestions that group dynamics and coherence are critical in risk perception, and even that people may in a sense choose perceptions in order to defend their group's way of life, is revolutionary in its ideas, deeply controversial, but still untested by convincing field evidence (Pidgeon et al., 1992).
Increasing understanding of the way in which risk perception is modified by psychological and social factors has accelerated developments in Risk Communication. The importance of this field is being increasingly recognized, due in part to the increase of legal requirements of the practitioners of risky activities to inform the public of risk levels, but also of the manifest political problems encountered by policy makers in the social tolerance of risk. The latter are amply illustrated by the widespread failure over the years of campaigns by authorities to convince the public of the acceptability of risks, and the analysis of the deep seated roots of outrage factors has illuminated the failings of the conventional approach of risk comparisons, for example, by comparing the risks of living near a nuclear reactor with that of walking across the road, to have much effect on public risk tolerance. Developments in this field are proceeding, and have so far few empirical studies to clarify issues currently only guessed at, but two developments seem clear (Pidgeon et al., 1992). First is that risk communication itself is currently an uncertain venture, which may "backfire" on its practitioners, exemplified in the cautionary conclusion that "one should no more release an untested communication than an untested product" (Warner, 1992). Second is that risk communication is becoming more of a two-way, interactive process, with the opinions of the public and of risk evaluators to be incorporated from the beginning of the risk analysis process, and constant feedback to be provided between assessors and evaluators in a "looped" process of analysis, to ensure the public relevance of the direction of investigation.
 
CONCLUSIONS
First, the traditional distinction between objective risk assessment and subjective risk evaluation is breaking down, for the two reasons given above.
The distinction drawn here between the roles of "assessment subjectivity", to assign probable damage levels in the face of incomplete data, and "evaluation subjectivity", for the consideration of the ethical or social significance of results, may clarify this problem to some extent but does not solve it. Nonetheless, the distinction remains a useful one, not least for the problems of the international harmonization of risk analysis procedures. At root, proposals for such harmonization entail the standardization and/or mutual recognition of risk assessment procedures, while leaving the value-laden risk evaluation process to sovereign states; yet the distinction is inadequately drawn in many national risk analysis processes, and the point on the chain of political responsibility where the line is drawn varies widely between countries.
Second, and contributing further to the problems of a clear assessment/evaluation distinction as a basis for harmonization, consensus is emerging that public perceptions of risk are best not treated as "problems" to be overcome but as valid inputs to the risk analysis process.
An important implication of this is that between nations with different risk perception cultures and traditions the entire process of risk analysis may be carried out in fundamentally different ways.
Third, the perception of risk across society varies between individuals in fundamental ways. As a result, the weighting of the risk perceptions of social groups, even when these can of themselves be adequately characterized, will always depend on political negotiation, compromise and the resolution of conflicts of perception.
"The interactive nature of ecosystems means that the testing of any selection of individual species will not gurantee that ecologically mediated effects will be detected."
 
3. ENVIRONMENTAL RISK ANALYSIS
THE TOXICOLOGY OF A SINGLE SPECIES
Typically, the assessment of toxic risk to a species follows a variant of the chain from hazard identification to risk evaluation discussed above. This typically takes the form, as exemplified here by the European Commission's Directive 93/67/EEC, "laying down the principles for the assessment of risks to man and the environment" of hazardous substances (European Commission, 1993), of: (1) Hazard identification, (2) Dose:response assessment, (3) Exposure assessment and (4) Risk characterization.
A. Hazard identification, the same in principle as the same process discussed in an engineering context above, here being "identification of the adverse effects which a substance has an inherent capacity to cause" (European Commission, 1993).
B. Dose: response assessment, carried out by the EC Principles to apply to both humans and "environmental compartments". A problem with this approach is the need for extrapolation, as discussed above. Another is the level of response to be assessed. The EC advocates, for risks to humans, the No Observed Adverse Effect Level (NOAEL) or, for environmental compartments, the No Observed Adverse Effect Concentration (NOAEC) or, where these are not feasible, the best or most appropriate equivalent such as applying an "assessment factor" to the LD50 or, if all else fails, using subjective or qualitative best guesses. The NOAEC is used to derive a Predicted No Effect Concentration (PNEC).
The use of this basic statistic is inevitable although it has difficulties. For example, at a conceptual level, the existence of "no effect" cannot be proved - the specification of no "observed" effect implies that unobserved ones may be present. Also, if dose:response relationships are continuous, a "threshold" may not exist, being a concept more of convenience than of science (Calow, 1993a, 1993b). Furthermore, the interval steps used for toxin concentrations in assays are sometimes large, and so NOAECs derived from them may be to some extent arbitrary (Barlow et al., 1992). The application of safety factors is intended, though it cannot guarantee, to remedy these difficulties, and thus the liberality of safety factors can be seen as a substitute for the precision of information.
The PNEC value is therefore in a sense itself a risk - the (low) risk of damage if a toxin is encountered at a certain concentration. It is thought to be zero, but this cannot be guaranteed: PNECs are "thought to be levels at which the probability of an effect is very low. This probability remains undefined; but the exercise of caution in the application of safety factors guarantees that it ought to be near zero" (Calow, 1993b).
C. Exposure assessment, whereby the likely exposure of susceptible components is assessed, whether by accidental releases or by conventional use of products such as cleaners, pesticides and paints, to obtain a Predicted Exposure Concentration (PEC). Exposure assessment, with a shorter history than toxicological dose:response assessment, is a science still in its infancy. It requires estimates of levels of production of the substance, patterns of its use, its distribution in the environment, including ways in which be transported in air or water, or bioaccumulated in organisms, and of its rates of degradation, as it may be broken down by sunlight, heat or organisms. Information about production and use may be available, and rates of bioaccumulation and degradation be researched by assays, but the remaining information is generally estimated by the use of mathematical computer models (Calow, 1993b). These are increasing in sophistication with, for example, the extension of the model past the periphery of the individual organism to the consideration of individual organs, and by kinetic models able to estimate the effects of pulsed and multiple doses and non-steady states (Landrum et al., 1992). However, they still require detailed calibration to be accepted (Calow, 1993a) and full quantification of the PEC is still unlikely. Regulators recognise that sometimes only qualitative estimates will be possible (European Commission, 1993). Like the PNEC, therefore, the PEC is also in a sense itself a risk - the risk of a certain concentration being encountered under envisaged conditions.
D. Risk characterization, the final step, is the estimation of the severity of likely effects. Its ideal is typically based on the risk that the PEC will exceed the NOAEC - most simply by the PEC/PNEC ratio (or "toxicity/exposure" or "T/E ratio"), to produce a convenient single figure by the socalled "quotient method". However, on the one hand the European Commission (1993) acknowledges that T/E ratios are only obtainable in the most accessible cases, being therefore an ideal which many analyses will fail to achieve, and on the other they are a rather simplistic expression of probabilistic information: as one observer commented on the use of the T/E ratio in the EC guidelines, "This is not quite risk assessment, in the sense of explicitly characterizing the probability of populations or communities becoming impaired to defined extents" (Calow, 1993a). In particular, the T/E ratio cannot meaningfully be a single figure, but should take account of the probability distributions of T and E: even if the average PEC is less than the PNEC, it may exceed it in certain areas if the toxin is unevenly distributed in the environment. Recommendation that these distributions be explicitly addressed is included in the Framework for Ecological Risk Assessment guidelines of the USA Environmental Protection Agency (USEPA) (Norton et al., 1992). Both the EC and USEPA guidelines emphasize the need for judgemental consideration of factors and information which are likely to be relevant in the interpretation of the "weight of evidence", but the USEPA is more precise in indicating that the details of T and E should be considered, such as their probability distributions and the possible influence of specific factors such as the vulnerability of species at different points in their life cycles. Final risk characterizations may, in view of the limitations of the data, be qualititative, by the T/E ratio, by estimation of the probability distribution of T and E or by more sophisticated presentations such as simulation models (Norton et al., 1992).
Some regulatory systems are essentially based only on toxicological data, the PNEC alone being used to inform the classification and labelling of products and so on. As the PNEC describes only properties of products, this approach is effectively the assessment of hazard, not of risk, and so may perhaps be better termed "hazard characterization", whereas the assessment of true risk by the comparison of PNECs with PECs, as required by the EC, has been called "risk assessment proper" - "specifying the likelihood of these effect concentrations being exceeded and hence having an effect on the target" (Calow, 1994).
 
E. The completion of risk characterization leads to recommendations for risk management. Decisions may be made in several ways: the ideal is by the acceptance of scientifically demonstrated quantified findings, but this is not always possible and, in fact, political evaluations often dictate conclusions as much as scientific findings. Policies may therefore be set by criteria of the cost of measures, by a process of political consensus or, indeed, by a more or less arbitrary decision of a standard which broadly satisfies political opinion - for example the commitment of industrialized countries, in Toronto in 1988, to reduce carbon dioxide emissions by 20% from 1988 levels by 2005 was largely made "regardless of the effectiveness of such a reduction" (Rotmans & Swart, 1990).
The conclusion reached may be, firstly, that the substance is too dangerous to be released or, secondly, that it may be released to the market provisionally and conditionally, until more information becomes available, possibly with risk management qualifications, such as limits on the total tonnage to be marketed or restrictions of use, such as its limitation to trained professionals, or stipulations for classification, labelling, packaging or the contents of the accompanying safety data sheet (European Commission, 1993; Greig-Smith,1992).
Many regulatory systems consider the total tonnage of products on the market, with assessment requirements incrementally increased as increasing tonnages marketed pass a series of threshold totals (European Commission, 1993).
Such management tools may be implemented in five basic ways, depending on the certainty of conclusions, the seriousness of errors and social and political factors. The first is by legal commands, such as statutory controls or bans, particularly appropriate when possible damage is serious and irreversible, such as potentially systematic risks. Second, policy may attempt to manipulate outcomes, such as by taxes or subsidies. Third, they may be broadly directed, a goal being stated and individuals being encouraged to meet it without specific statutory commands, such as by recommendations to use "Best Practical Means", the "Best Available Technology Not Entailing Excessive Cost", and so on. Fourth, there may be requirements for information, such as the provision of data to the public. Fifth, processes may be specified, such as public enquiry structures to reconcile conflicts in specific cases.
In general, the trend in many countries is of a move away from recommendations based solely on hazard characterization, such as attempts to influence use so that the PNEC is not exceeded, such as by specifying various "Best Available Technology" practices, or simply to require that it be not exceeded. Instead, there is a move to the use of EC-style risk assessments proper, taking into account the patterns of use, which allows management by other means, such as limiting the total tonnage of product released to the market (Calow, 1993b).
On the other hand, the conclusion may be that the analysis is not yet over, and that more information is required, to be obtained by further laboratory tests or by field trials, in which case, as funds for such research are not unlimited, objectives must be prioritized by their likely cost-effectiveness (Calow, 1993a).
 
F. Researchers in the field are in agreement over the importance, after a decision has taken effect (such as the evaluated product being released for sale), of post facto monitoring of impact to evaluate the risks and risk analysis procedures involved, but regulatory requirements have been slow to incorporate this. Such monitoring may bring several benefits, both for the risk management of the case in question and for the gathering of information of wider use to the development of the field as a whole.
First, failures of the risk evaluation may be rectified, and conclusions revised, such as by the withdrawal of a product; second, predictions can be checked; third, the capacity of environmental systems for recovery from impact may be evaluated; fourth, unforeseen ecologically mediated effects may be evaluated; and, fifth, information may be gained to assist the all-important understanding of processes. The ideal would be a continuous, rolling process of checking and evaluation, continually updating the status of knowledge by, for example, the incorporation of post-release findings, whether from similar products or different cases and countries, into the base of risk information, to be integrated by methods such as Bayesian techniques for the ongoing refinement of the understanding of risks. For incremental risks, indeed, post-release monitoring may to some extent draw the sting from the entire risk management process: if errors in the initial risk analysis can be made good, then the risk of irreversible catastrophic failure is effectively removed. Here, indeed, may lie a danger of permitting a false sense of security - errors may not be rectifiable in the cases of human suffering or of systematic and irreversible damage to ecosystems, and indeed this principle is the whole point of risk analysis in the first place.
 
ECOSYSTEM TOXICOLOGY
In the light of the difficulties of the description of ecosystems due to their inherent complexity, the tangled web of ecologically mediated effects such as competition, predation and bioaccumulation, and of the variety of ways in which they may be considered to be important or significant, their effective risk analysis depends on the choice of "endpoints" (Lipton et al., 1993). These are the specific criteria which are to be assessed for damage by the stressor in question, chosen to embody or to represent the ecosystem features which are considered to be of value. A panel of experts surveyed by a Delphi process was in agreement that good endpoints, which may be considered as hypotheses to be tested, should have accurate measurability, utility, relevance, importance in the ecosystem (however defined) and, ideally, a body of accompanying information such as toxicological and ecological background knowledge and the possibility of controls such as measurements on comparable but undisturbed sites. It also considered that endpoints are rarely clearly enough defined and that, even when they are, the logistic arrangements for their measurement, particularly in the long term, are underestimated and inadequate, such as in the use of too small samples (Richey et al., 1985b).
There is a danger that endpoints be applied to the symptoms of ecological damage, rather than to their fundamental causes, by damage to functions. Symptoms are by their nature more easily detectable, and so attention may be focused on them. However, less visible damage to functions, such as the oxidising capacity of the atmosphere, may be of critical importance, particularly if it is not identified in time for remedial action to be taken.
Endpoints can be divided into the final criteria to be the objects of the assessment ("assessment endpoints") and their more readily measurable proxies which are actually assessed ("measurement endpoints") (Suter, 1990). The assessment endpoint is a "formal expression of the actual environmental value to be protected". As qualities, they should have social and biological relevance, unambiguous operational definition and susceptibility to the stressor being assessed. Examples include recreational value, the size of an important population, biological diversity, beauty, soil stability or the resistance of the system to outbreaks of pests or fires. The selection of assessment endpoints is essentially subjective and political - individuals of charismatic or endangered species may be considered more important than entire populations of small and unexciting invertebrates. Species may be given relative importance in different ways, such as a simple checklist, a ranking or weighted scores: these may also be combined, for example to protect a few important species, and then maximize the diversity of all the others.
Sometimes balanced decisions must be made, such as one recent evaluation which explicitly gave greater importance to populations of (rarer) geese than those of (more abundant) pigeons (Greig-Smith, 1992).
Measurement endpoints should be easily, quickly and cheaply measurable and clearly related both to the operation of the stressor and to the assessment endpoints they are intended to represent, preferably in a quantifiable way. Examples include estimates of population density, counts of species or diversity indices, quantified descriptors of landscape quality, assessments of soil loss in water runoff, and the frequency of pest or fire outbreaks (Suter, 1990).
 
A. THE ASSESSMENT OF ECOSYSTEMS AS REPRESENTED BY INDIVIDUAL SPECIES
Clearly, not every species in an ecosystem can be evaluated for vulnerability. As a result, certain particular species are chosen as measurement endpoints. One view of the selection of these is that critically vulnerable species can serve as "sentinel species", a substance found harmless to which can be presumed to be harmless to all the others.
It has been argued that single-species tests of this sort can be reliable if the species is carefully and skillfully chosen, but the risks of such a choice themselves make the use of a suite of species more attractive - species vary in their sensitivities to different toxins, so that one sensitive to one stressor class may be relatively resistant to another (Cairns, 1989).
Test species can be chosen by several criteria, principally as being particularly sensitive, important, whether to the structure or function of an ecosystem or, economically or aesthetically, to humans, representative of a taxonomic or trophic group or convenient, being suitable to be reared, housed and tested in adequate numbers. Most test species are chosen for the last of these criteria, convenience, and efforts are being made to enhance the reliability of extrapolation of data from a handful of convenient species to those likelier to be selected as assessment endpoints. Progress is being made in this field. For example, although most easily laboratory-rearable species such as Daphnia magna tend to be rather insensitive to toxins, being fecund, robust, ubiquitous and herbivorous (herbivores, tending to encounter natural poisons in their plant food, tend to be more resilient to toxins than carnivores), not all are, and the use of relatively sensitive test organisms, such as oyster larvae, may be expected to increase (Gray, 1989). Also, evidence is emerging that the mechanisms of slow-acting chronic toxicity, which is by its nature more time-consuming to assess than acute toxicity, are, luckily, in general more widespread than those of acute, and thus that chronic toxicity data may be more easily extrapolable (Barlow et al., 1992).
Attempts are also being made to introduce some meaningful harmonization of tested species, either by loose classifications, such as "a typical 25g seed-eating bird" (Greig-Smith, 1992), or by aggre-gation to form categories such as "the biomass of bottom-feeding fish" (Richey et al., 1985).
In general, a pattern is emerging of a process of detailed, highly replicated and expensive tests in research laboratories supporting the reliability and extrapolability of cheaper and simpler tests carried out in testing laboratories. Integration of these two activities may enable a coherent process to emerge - test laboratories using a few easily rearable and well understood test organisms, and detailed research results being used to draw inferences about extrapolation to assessment endpoints. For example, testing for damage done by a stressor, when the test organism does not actually die, is in fact not easy, no single technique being easily selectable. General tests for toxic damage do exist, taking advantage of the fact that energy is needed to resist any stress, to measure a stress by looking at cellular respiration rates and other basic processes such as protein synthesis. Unfortunately, such elegant and versatile techniques are difficult and expensive, and so not widely used in testing, but they may be used by research laboratories for studies of extrapolation. Also the tools developed over decades for extrapolation of toxicological data to humans are now generally well established and internationally recognised, and may be used as models for similar extrapolations to other species (Barlow et al., 1992).
This emphasis on extrapolation is ultimately inevitable, in the context of international harmonization, to resolve the fundamental opposition of the needs of harmonization and relevance. Ecosystems vary - a chemical in the environment in, say, Canada may well encounter none of the same species in Indonesia - and so the likely responses of indigenous species must be addressed in risk characterization if the process is to have any relevance. If, therefore, some extrapolation is essential if assays are not to be duplicated in virtually every country where a product may be launched, the test species may as well be from a relatively small selection, as long as the toxicology of the species is well understood and the assessments are controlled by ring tests, and detailed research results may then be used to inform the extrapolation of these findings for specific national ecosystems.
 
B. THE HOLISTIC ASSESSMENT OF ECOSYSTEMS
The interactive nature of ecosystems means that the testing of any selection of individual species will not guarantee that ecologically mediated effects will be detected. Again, for more holistic analyses of ecosystems, endpoints must be carefully chosen and formally stated because, as outlined above, the value attached by society to an ecosystem may depend on its diversity, functions or the presence of glamorous species.
Endpoints chosen may be aspects of ecosystem structure (which species are present and their abundance), such as total diversity or the presence of rare or attractive species, or ecosystem function (the processes which it carries out), such as carbon dioxide uptake. The former are generally chosen, as the maintenance of the structure of an ecosystem will usually maintain its functions, but not vice versa. In other ways endpoints may overlap and coincide: many glamorous species, with high political, social and cultural values, such as the tiger or the monkey-eating eagle (a national symbol of the Philippines), are "top predators" at the summit of food chains, roving over large areas, and as a result a viable population of individuals requires a huge area of habitat - so the protection of such a species necessarily entails the protection of such areas of habitat, with the corollary of simultaneously protecting its other plant and animal species.
Logically, the extension of toxicological risk analysis from individuals to ecosystems is simply one more step up a chain of levels of organisation; from biochemical tests to tests in vitro, and thence on to organs, to organisms and so to ecosystems (Barlow et al., 1992). However, they add more than one incremental level of complexity, entailing the addition of wildlife ecology and ecological toxicology to the scientific disciplines of analytical and biochemical toxicology etcetera used for lower levels of organization (Kendall & Akerman, 1992). Above all, there is the possible existence of "risk cascades", in which human disturbances hoped to be incremental turn out to be systematic, and ecological interactions magnify the effects of damage in reverberations through the whole system (Lipton et al., 1993).
The most obvious tool which can be employed in the light of these difficulties is experimental toxicological tests on multispecies systems.
Such systems are notoriously difficult to maintain in equilibrium even when undisturbed, as in closed systems of any manageable size predators tend to eat all their prey and then starve, but much good progress has been made in the design of these systems, by unglamorous but valuable "bread and butter" research to evaluate how manageably small systems may be stabilized: for example, ways have recently been described in which woodlice and bacteria in a laboratory system together decompose leaf litter, an important ecological process in deciduous forests (van Wensem, 1989), and in which predator and prey fishes may be maintained together in smaller tanks than was previously thought possible, if vegetation provides an environment of sufficient structural complexity to allow the prey some refuge to prevent them from all being eaten (Harick et al., 1993). Toxicological tests can be carried out in such systems, and some regulatory authorities do now demand multispecies testing for the registration process of toxins.
Another possibility is presented by the fact that, unlike in human toxicology, limited studies in the field can be carried out - single streams or ponds, whether artificial or natural, for example, may be experimentally contaminated, or otherwise disturbed, and the results assessed without risk of serious damage to the wider environment. Such experiments are particularly important, because the data they provide can be used for the calibration of laboratory test results, for the extrapolation of these results to estimate likely effects in the field. At the moment, work is proceeding on the issues needed for this extrapolation, including some very detailed and long-term studies, with some success, though more field tests are needed for reliability (Emans et al., 1993).
This work is important to the goal of the USEPA and other authorities to moving towards holistic whole-ecosystem evaluation, away from a "pollutant-by-pollutant and medium-by-medium" approach (Bretthauer, 1992).
Ultimately, however, not all ecosystem risks can be quantified by tests prior to release. Subtle effects on wildlife, such as behavioural changes, for example reductions in success in courting a mate, caring for young or avoiding predators, may only slowly become apparent. Three activities are important as a result - the continual monitoring of effects after release, the development of computer models of ecosystems and the use of imaginative "what if?" examinations, using expert judgement, to attempt to consider what unforeseen complications may arise.
Post-release monitoring of ecosystems, like that of individual species, is useful not only to look for failures in the specific release in question, but for general lessons for use in future: for example, the survival of the species originally used for laboratory tests may provide information about its usefulness as a test species, and the validity of extrapolation processes can be assessed with the benefit of hindsight.
Ideally, a monitoring programme should be fully thought out in advance, as part of the original risk assessment design, so that the information it provides is directly relevant to the rest of the process. A comprehensive monitoring programme would include initial surveys of species present and their toxicological status, including natural levels of variation against which the effects of disturbance may be assessed, as well as subsequent recording of all status changes, including those initially not apparently attributable to human disturbance. At present, such procedures are not much internationally harmonized, and they may benefit from a coordinated approach, for example by the extension of systems such as Britain's River InVertebrate Prediction And Classification System (RIVPACS), a checklist of species expected to be found in pristine rivers, ranked in order of sensitivity to pollution, so that any particular water body can be awarded a value for ecological health by which species are present (Barlow et al., 1992).
Mathematical and computer models of ecosystems, allowing the quantification of relationships between species, are difficult to build and test, and have indeed been an objective of ecological scientists for several decades. However, some progress is being made, such as in energy budget models, which model the flows of energy as nutrients through food chains, and which can be used to assess complex outcomes such as the impact on the viability of a population of delays in its reproduction cycle, and with further work their use should bring considerable benefits (Barlow et al., 1992).
For the time being, much reliance is placed on subjective expert consideration of likely ecological effects. Here too there is scope for the formalization, and thereby harmonization, of procedures. For example, an ecosystem may be considered as, in effect, a huge three-dimensional table, the three axes being species, effects and assessment purposes. Each cell of the table may be addressed in turn, consideration being given to the appropriateness of the available information and the adequacy of the risk characterization, thus giving systematic structure to the consideration of failings in the holistic picture (in a manner rather reminiscent of the use of hazard identification procedures such as GENHAZ).
Such procedures may also help the formalization of the documentation trail, and the development of consistency in their use would benefit the clarity and cross-compatibility of risk analyses (Greig-Smith, 1992).
 
NONTOXICOLOGICAL AND SYSTEMATIC STRESSES
The nature of nontoxicological and systematic stresses ensures that their risk analysis must be more flexible and specific than those of toxins.
This is because they come in a wider array of types, requiring flexibility, and tend to be site-specific, applicable to processes, not products, from activities such as timber cutting and civil engineering projects such as water diversion, roads or airports. Whether or not factors such as "percentage deforestation" can be treated as analogous to toxic doses, as proposed by the USEPA (Norton et al., 1992) is still in some doubt, and may well not be the case.
The setting of endpoints and the quantification of responses to disturbance, on the other hand, are broadly analogous to those in toxicological assessments. For example, disturbances to Alaskan Caribou by traffic around oil field developments have been assessed by simple but illustrative "activity budgets" which evaluate changes in the time animals spend on various activities, such as eating, in response to passing vehicles (Murphy & Curatolo, 1987); and even an endpoint as elusive as landscape beauty can be both quantified, by landscape descriptors, and rated for value (and therefore the gravity of damage be assessed), by either hedonic or contingent valuation techniques (Willis & Garrod, 1993).
Risks of systematic disturbance to ecosystems, being by nature imponderable, specific and difficult to test in advance, rely particularly heavily on expert estimation of likely impacts, and expert selection of the best approach to the analysis. As a result, generalizations about the specific techniques which may be used is hard, and the issues raised tend to be less technical than those discussed above, and more in terms of the sociopolitical and institutional frameworks in which they are considered and discussed. This can be seen from an example of the analysis of a systematic risk, the proposed introduction of an exotic fish, the channel catfish, to New Zealand.
New Zealand, as an "ecological island" whose aboriginal wildlife has already been ravaged by the injudicious introduction of exotic species in the past, is sensitive to their risks. The proposed catfish introduction was not to the wild, but for fish farming, and an initial Environmental Impact Assessment (EIA), including consideration of the consequences of escapes to the wild, suggested the risks would be acceptable. Because of public concern, however, two independent experts were mandated to assess all the available existing information. Largely as a result of efforts by a pressure group opposed to the introduction (themselves, as it happens, not defenders of New Zealand's precarious aboriginal wildlife, but sport anglers concerned at possible impacts on their prey populations of trout introduced decades earlier), new information came to light which had not been considered before. This included documented cases of catfish escapes to the wild from farms, and reports from fish ecologists, in areas where the catfish had been introduced, of its rapid spread into a variety of habitats and of damage through predation and competition to many indigenous species, including some similar in many respects to endangered species in New Zealand. Reviewing the lessons to be learnt from their final, accepted, recommendation that the introduction was unacceptably risky and should not be carried out, the independent experts drew several conclusions. One was that an independent review should be the last stage before release. Another, more important, was of the role of publicity in uncovering the maximum possible amount of relevant information: as they put it, "It is not enough to expect that the compiler of the EIA will have uncovered all salient facts about the proposed species". In this case, much important scientific information was obtained from areas where introductions had been studied, often from obscure, local journals, and even as personal communications from scientists of findings which had never been published. The publication of requests for information relevant to an EIA, particularly through specialist organs such as professional associations and journals, is a useful proposal to maximize the information obtained (Townsend & Winterbourn, 1992).
 
CONCLUSIONS:
Paradigms for Ecological Risk Assessment
Attempts are being continued to establish a general paradigm for ecological risk assessment: the paradigm for human environmental health, of the basic sequence of hazard identification, dose:response estimation, exposure assessment and risk characterization, has achieved a wide consensus, but is not strictly or directly applicable to ecosystem risk analysis, with its multitudes of stressors, species, orders of organization and interactions between them. One recent suggestion (Lipton et al., 1993) has been for the following structure:
1). Receptor identification of species and functions of concern;
2). Hazard identification;
3). Endpoint identification, of the response of which receptors to which stressors are to assessed;
4). Relationship assessment, for the consideration of ecologically mediated effects and the possibility of risk cascades, the results of which may, if necessary, be fed back to (2) as the identification of new hazards;
5). Exposure assessment;
6). Response assessment (renamed, in view of the imperfect quantifiability of many ecological stresses, from "dose:response assessment");
7). Risk characterization and uncertainty analysis.
Paradigms such as this are still in development. A feature which they share, however, is allowance for feedback of information in loops, so that earlier points of the process may be returned to for reevaluation in the light of later findings, and mutual communication between the evaluator and assessor to ensure the political relevance of the factors assessed (Norton et al., 1992).
4. CONCLUSION
The field of environmental risk analysis is developing fast: changes are currently under way, and more may be expected in the near future. Changes are driven by three main motors.
First is the growing demand for it, thanks to continual growth both of public awareness of environmental issues, and of the organization and knowledgeability of environmentalist public opinion, raising public expectations of the quality of environmental risk analyses and demanding the publicization of risk analyses for informed public debate. Increasingly the public is not prepared to leave the discussion of environmental risks to "experts", but is coming to expect their debate in more open fora. The production of skilled, experienced and articulate environmental specialists currently seems to be lagging behind this demand.
Second is a growing search for ways to formalize and to rationalize environmental risk analyses. In spite of the variability, uncertainty and value judgements encountered in environmental analysis (or often, in fact, because of them), it is becoming felt that they cannot be left to woolly and subjective techniques relying heavily on largely intuitive expert judgement. The result is a search both for tools which can rationalize the application of expert judgements, and for logical and formal structures and frameworks within which risk analyses can be carried out.
Third is the desire, driven by attempts to promote international trade, whether on regional (NAFTA; the European Community) or global (GATT) scales, to harmonize environmental risk analysis procedures, in ways which retain the rights of sovereign states to set their own environmental policies and priorities and yet do not needlessly inhibit international commerce. Ideally, technical issues such as procedures, tests, the types of data needed and the manner of their collection could be harmonized, by international agreement, while leaving the level of acceptable risk to be set by sovereign states.
These three developments have many qualities in common. Harmonization and rationalization/formalization of procedures clearly go hand in hand, the former requiring the latter, and in many instances will both reduce the workload of individual experts (such as by the use of expert systems) and increase the transparency of the analysis process.
Harmonization of environmental risk analysis procedures therefore brings four advantages. First are the economies of scale from mutual recognition of processes and results, removing the need to duplicate in each country tests already performed elsewhere. Second are the benefits of the enhancement of trade. Third is the impetus it adds to the ongoing process of the rationalization of procedures, the development, refinement and evaluation of techniques, and a fruitful debate, in panels, workshops and publications, of what environmental risk analysis is for and how it may be best fitted to our purposes. Fourth is its contribution to the development of common databases and pools of information and knowledge. At a basic, technical level many of these are well advanced, such as the EC "black" and UK "red" lists of chemicals with data on their toxicity, persistence and capacity for bioaccumulation. Higher level databases can also be developed, but require flexibility in their construction, to allow for the inclusion of estimates of the uncertainty of information, detailed background information such as the genotypes, rearing conditions and diets of organisms tested, and comments and opinions by the scientists involved.
At the technical level, much progress has been made over the last two decades, such as laboratory ring tests and the OECD's recommendation of standard tests, good laboratory practice guidelines and test protocols. The European and Mediterranean Plant Protection Organization and Council of Europe scheme for the risk assessment of plant protection products is another example (Barlow et al., 1992).
Another development is the arrival at consensus over the importance of information and its communication, in five interconnecting ways.
First is the importance of the documentation of analyses, as full "documentation trails", so that individual analyses can be case-specific but their components assessed, per se and for possible use elsewhere. These should include the subjectively estimated components, with rationales and justifications of why decisions were taken the way they were - these decisions cannot be understood without some documentation of the reasoning processes used, and of the background knowledge and awareness of the experts participating, as the New Zealand catfish example shows.
Second are the implications of the widening understanding of the fallibility of the distinction between "objective" risk assessment and "subjective" risk evaluation, as understanding advances both of the deep-seated nature of public risk perceptions and of the limits to the objectivity of expert scientists. The operational conclusion of this is recommendations that the information flow in risk analysis should be a two-way process of information feedback, both between the risk evaluator and the scientific assessors (Norton et al., 1992) and between the public and the risk analysis team (Pidgeon et al., 1992), so that the criteria of final subjective acceptability can be recognised and incorporated throughout the analysis process from its beginning.
Third is recognition that for meaningful evaluation, the presentation of scientific results of risk assessment should include estimates, and discussion of the possibility, of uncertainty about values and estimates, so that each datum is presented with a "pedigree" allowing the evaluator to assess the confidence held in it. The results of uncertainty and sensitivity analyses can usefully also be included in this context.
Fourth, the usefulness of post facto evaluation, validation and monitoring is widely acknowledged. Its usefulness for the all-important development of the understanding of mechanisms and for the calibration of extrapolation would be enhanced if it were more widespread than at present, particularly if done in a standardized, harmonized way so that a large body of knowledge from validation monitoring could be built up (Greig-Smith, 1992), and if its findings were accessible with the documentation trail of the original analysis, so that lessons could be learnt from the consideration of the two together.
Fifth, the importance of political and institutional processes is a recurrent theme. Risk analyses proceed better in institutional environments which allow the gathering and communication of information to be maximized.
This principle extends to the discussion of risk in the public sphere, where the trustworthiness of institutions, and their openness to the reception of public views and effectiveness in communication are important, and increasing requirements are placed on operators of risky facilities to inform the public of their natures.
On the other hand, several aspects of the rationalization and harmonization of environmental risk analysis procedures are contentious.
There are principally three issues to be balanced against their benefits.
First is the question of differences in cost and technical sophistication between techniques. The most expensive and sophisticated tools may not always be considered cost-effective, or even be available to poorer countries, particularly if they rely on patented processes held elsewhere. As assessment technology develops particularly quickly, some geographical and other areas will inevitably move faster than others in the adoption of new techniques.
Second is the importance of flexibility in analysis procedures to allow the precise tailoring of an analysis to its circumstances. This is particularly true of the geographically specific field of Environmental Impact Analysis, which applies chiefly to process rather than product standards, but these too are increasingly under discussion in the trade and the environment debate. Product risk analyses also need tailoring to specific cases, as ecosystems and their vulnerability vary throughout the world, and so does the capacity of environmental processes to degrade, immobilize or disperse contaminants.
Third is the inevitable issue of differences in the values which societies attach to environmental risks - how they rank the environment, development, employment, mobility and other economic and social goals. The frailty of the distinction between risk assessment and evaluation means that it is not easy to distinguish, as free trade interests would hope, that risk assessment can be internationally harmonized on the sole criterion of "sound science" and evaluation left to sovereign states (Morris, 1993).
The ways in which these three criteria conflict and interact with the aims of harmonization are manifold. They may be considered further by a look at the areas in which the objectives come into conflict (Greig-Smith, 1992).
1) The choice of a test protocol, whether to use extrapolation, modelling, a standard test or a test chosen for its appropriateness. Here a standard protocol will enhance compatibility and the evolution of a usefully comprehensive knowledge base. On the other hand, flexibility of choice allows for selection to suit particular conditions, the use of information from a wider variety of sources and the rapid adoption of newer techniques as they become available. In particular, the relative advantages of these four approaches may not remain the same: models are inherently holistic, and do not rely on a large supporting base of information, but may not yet be empirically validated, and so their usefulness may grow. A possible compromise may be to advocate the use of standard procedures but to permit deviation from them when good reasons can be stated and justified.
2) When a test protocol is chosen, how tightly it may be standardized. This is shown by the question of whether in bioassays to use clones of test organisms such as Daphnia magna, or to use a more diverse spread of genotypes by, for example, taking them from the wild. The use of standard, identified clones is the only way to ensure compatibility of results. On the other hand, clone populations may poorly predict the responses of wild populations - lacking genetic diversity, they may be peculiarly susceptible to certain stressors (for example, by lacking a resistance mechanism which is in fact widespread in the wild), and they may simply not be representative of wild populations, particularly on an international scale.
The only solution to this problem is to search for reliable extrapolation techniques so that clones may be used for tests internationally, and the results of detailed research used for extrapolation from the internationally standardized findings to the likely situation in any particular national case.
3) Whether to take a balanced view of risk-taking, or a safety-first "presumption of hazard" approach, with the burden of proof resting on those who would benefit from the taking of the risk. This decision is not the sole concern of risk evaluation, but permeates risk analysis, as the choice made will affect issues such as experimental design and sample sizes.
Presumption of hazard is cautious and reassuring, but may be overcautious, and has scientific difficulties in the impossibility of proof of "no effect". It may be possible to combine these approaches to some extent by, for example, varying the value of delta in tests of "environmental significance".
4) Whether or not to use "trigger" values in testing, to simplify the decision to advance the procedure to the next step. Trigger values enhance harmonization and reduce the need for unnecessary testing, but are less flexible and realistic than variable levels.
5) The flexibility of safety factors. Rigid factors enhance confidence, objectivity and harmonization, but flexible ones reduce the risks of overcaution and of loss of information, and can be modified in consideration of, for example, the importance of the species in question or the uncertainty of the information.
6) The use of expert opinion. To use expert opinion maximizes flexibility and the use of information, but ultimately cannot be fully consistent as opinions are bound to differ in some ways.
7) Whether to summarize risks by categorization as "high", "medium" or "low", or to attempt to present precise values. Scoring systems, which are growing in use (Calow, 1994), may be seen as intermediate between the two.
Categorization allows for some basic harmonization and recognizes the uncertain nature of risk information - gross classes may best reflect uncertainty. Yet it may underuse information, and the category definitions themselves are subjective and imprecise. Procedures can be used, however, to allocate results to categorization classes and then to extend research when, and only when, findings are felt to be near the boundary between two classes, so adding precision only when necessary.
8) Whether to use similar categorization for levels of uncertainty in estimates, with similar arguments for and against.
9) Whether to base risk analysis and management essentially on hazard characterization, by recommendations intended to ensure that Predicted No Effect Concentrations are not exceeded, or whether to attempt full risk assessment by the consideration of Predicted Environmental Concentrations as well. In fact, for international harmonization purposes, this need not be important, as PNECs may be expected to be more widely applicable than PECs, the latter varying more widely between countries with differences in levels of industrialization, agricultural practices and so on, and so it may be possible to proceed towards harmonization of PNEC procedures and data, while leaving PEC assessment to national authorities.
10) Whether to regard environmental risks as absolutely to be avoided, or explicitly to balance them against the benefits of taking risks, by a comparison such as Risk:Benefit Analysis. Again, it may be possible to standardize estimations of risk, while leaving the questions of whether to balance them against benefits and, if so, what benefit level may justify risks, to sovereign states.
It can be seen from this list that precise and universal regulation of the issues surrounding harmonization may not be possible, but also that integration of more than one approach can sometimes be done.
Broadly speaking, it may be possible to harmonize the entire structure of environmental risk analysis in two ways. The first would be to maintain or clarify the distinction between objective assessment and subjective evaluation, on the assumption that the scientific groundings of assessments may be universally recognised, and the evaluative phase added on as a superstructure by sovereign states. Despite differences in basic evaluations of the significance of risk assessment, this may be possible, at least between countries with similar environmental priorities (such as those within the European Community). The second would be to leave all subjective and evaluative components in risk analysis open to modification, by full documentation of the decision processes at every point of an analysis. Many practitioners in the field argue that this is desirable anyway, for the establishment of a knowledge base and the transparency of the process to risk evaluators and to the public. These two outlines are not mutually incompatible and may be pursued together.
Below the grand structural framework of analyses there is an enormous amount which can be and is being done in the formalization and harmonization of tests, decision-making procedures and the use and documentation of expert judgement. These developments are not only driven by international issues, nor do they contribute solely to them, but also derive intrinsically from the progress of environmental risk analysis itself as it matures into a coherent field of scientific endeavour.
--------------------------------------------------------------------------
REFERENCES
Baird, D.J., Barber, I., Bradley, M., Calow, P., & Soares, A.M.V.M. (1989). "The Daphnia bioassay: a critique." Munawar et al., 403-6.
Barde, JP., & Pearce, D.W. (eds.) (1991). Valuing the Environment: Six Case Studies. London,
UK: Earthscan Publications. ix+27pp.
Barlow, S.M., Bridges, J.W., Calow, P., Connong, D.M., Curnow, R.N., Dayan, A.D., Purchase, I.F.H., & Suckling, C.W. (1992). "Toxicity, toxicology and nutrition." Royal Society, 35-65.
Berry, C.L. (1990). "The hazards of healthy living - the agricultural component." British Crop Protection Council Conference - Pests & Diseases 1, 3-13.
Blalock, H.M.jnr (1961). Causal Inferences in Nonexperimental Research. Chapel Hill, NC, USA: University of North Carolina Press. xii+200pp.
Bretthauer, E.W. (1992). Editorial: "The challenge of ecological risk assessment." Environmental Toxicology & Chemistry 11, 1661-2.
Cairns, J.jnr (1989). "Foreword." Munawar et al., xiii-iv.
Calabrese, E.J., & Baldwin, L.A. (1993). Performing Ecological Risk Assessments. Chelsea, MI, USA: Lewis Publishers. xxiii+257pp.
Calow, P. (1992). "Can ecosystems be healthy? Critical consideration of concepts." Journal of Aquatic Ecosystem Health 1, 1-5.
Calow, P. (1993a). Editorial: "Hazards and risks in Europe: challenges for ecotoxicology." Environmental Toxicology & Chemistry 12, 1519-20.
Calow, P. (1993b). "Hazarding an estimate of risk." Integrated Environmental Management 20 (June), 2-5.
Calow, P. (1994). "Risk Business." Integrated Environmental Management 27 (March), 10-11.
Charnowitz, S. (1993). "Environmental harmonization and trade policy." Zaelke et al., 267-86.
Crossland, B., Bennett, P.A., Ellis, A.F., Farmer, F.R., Gittus, J., Godfrey, P.S., Hambly, E.C., Kletz, T.A., & Lees, F.P. (1992). "Estimating engineering risk." Royal Society, 13-34.
Dillon, J.L., & Officer, R.R. (1971). "Economic versus statistical significance in agricultural research and extension: a ProBayesian view." Farm Economist 12, 31-45.
Emans, H.J.B., van der Plassche, E.J., Canton, J.H., Okkerman, P.C., & Sparenburg, P.M. (1993). "Validation on some extrapolation methods used for effect assessment." Environmental Toxicology & Chemistry 12, 2139-56.
European Commission (1993). "Laying down the principles for assessment of risks to man and the environment of substances notified in accordance with Council Directive 67/548/EEC." Official Journal of the European Communities L227(8.9.93), 9-18.
Geraghty, P.J. (1993). "Environmental assessment and the application of expert systems: an overview." Journal of Environmental Management 39, 27-38.
Giesy, J.P., & Graney, R.L. (1989). "Recent developments in and intercomparisons of acute and chronic bioassays and bioindicators." Munawar et al., 21-60.
Gray, J.S. (1989). "Do bioassays adequately predict ecological effects of pollutants?" Munawar et al., 397-402.
Greig-Smith, P.W. (1992). "A European perspective on ecological risk assessment, illustrated by pesticide registration procedures in the United Kingdom. Environmental Toxicology & Chemistry 11, 1673-89.
Harick, G.L., deNoyelles, F., Dewey, S.L., Mason, L., & Baker, D. (1993). "The feasibility of stocking largemouth bass in 0.04ha mesocosms used for pesticide research." Environmental Toxicology & Chemistry 12, 1883-93.
Johnston, G., Walker, C.H., & Dawson, A. (1994). "Interactive effects of Prochloraz and Malathion in pigeon, starling and hybrid red-legged partridge." Environmental Toxicology & Chemistry 13, 115-20.
Kendall, R.J., & Akerman, J. (1992). "Terrestrial wildlife exposed to agrochemicals: an ecological risk assessment." Environmental Toxicology & Chemistry 11, 1727-49.
Landrum, P.F., Lee, H.II, & Lydy, M.J. (1992). "Toxicokinetics in aquatic systems: model comparisons and use in hazard assessment." Environmental Toxicology & Chemistry 11, 1707-25.
Lipton, J., Galbraith, H., Burger, H., & Wartenberg, D. (1993). "A paradigm for ecological risk assessment." Environmental Management 17, 1-5.
Maki, A.W. (1992). "Of measured risks: the environmental impacts of the Prudhoe Bay, Alaska, oil field." Environmental Toxicology & Chemistry 1,1 1691-707.
Morris, R.J. (1993). "A business perspective on trade and the environment." Zaelke et al., 121-32.
Munawar, M., Dixon, G., Mayfield, C.I., Reynoldson, T., & Sadar, M.H. (eds..) (1989). Environmental Bioassay Techniques and their Application. Hydrobiologia 188/9 (special volume). xiv+680pp.
Munawar, M., Munawar, I.F., Mayfield, C.I., & McCarthy, L.H. (1989). "Probing ecosystem health: a multi-disciplinary and multi-trophic assay strategy." Munawar et al., 93-116.
Murphy, S.M., & Curatolo, J.A. (1987). "Activity budgets and movement rates of caribou encountering pipelines, roads and traffic in northern Alaska." Canadian Journal of Zoology 65, 2483-90.
Norton, S.B., Rodier, D.J., Gentile, J.H., van der Schalie, W.H., Wood, W.P., & Slimak, M.W. (1992). "A framework for ecological risk assessment at the EPA." Environmental Toxicology & Chemistry 11, 1663-72.
Pidgeon, N., Hood, C., Jones, D., Turner, B., Gibson, R. (1992). "Risk perception." Royal Society, 89-134.
Richey, J.S., Mar, B.W., & Horner, R.R. (1985). "The Delphi technique in environmental assessment: 1. Implementation and effectiveness." Journal of Environmental Management 21, 135-46.
Richey, J.S., Horner, R.R., & Mar, B.W. (1985). "The Delphi technique in environmental assessment: 2. Consensus on critical issues in environmental monitoring program design." Journal of Environmental Management 21, 147-59.
Rotmans, J., & Swart, R. (1990). "The gloomy greenhouse: should the world phase out fossil fuels?" Environmental Management 14, 291-6.
Royal Commission on Environmental Pollution (1991). GENHAZ: A System for the Critical Appraisal of Proposals to Release Genetically Modified Organisms into the Environment. London, UK: HMSO. vii+55pp.
Royal Society (1992). Risk: Analysis, Perception and Management. London, UK: The Royal Society. vi+201pp.
Seager, J., & Maltby, L. (1989). "Assessing the impact of episodic pollution." Munawar et al., 633-40.
Siegel, S., & Castellan, N.J.jnr (1988). Nonparametric Statistics for the Behavioral Sciences. Second Edition. New York NY, USA: McGraw-Hill. xxiii+399pp.
Slovic, P. (1993). "Perceived risk, trust and democracy." Risk Analysis 13, 675-82.
Spalding, H., & Smit, B. (1993). "Cumulative environmental change: conceptual frameworks, evaluation approaches and institutional perspectives." Environmental Management 17, 587-600.
Stewart, A. (1993). "Environmental risk assessment: the divergent methodologies of economists, lawyers and scientists." Environmental Law & Planning Journal 10, 10-8.
Suter, G.W.II (1990) "Endpoints for regional ecological risk assessments." Environmental Management 14, 9-23.
Talcott, F.W. (1992). "How certain is that environmental risk estimate?" Resources Spring, 10-5.
Townsend, C.R., & Winterbourn, M.J. (1992). "Assessment of the environmental risk posed by an exotic fish: the proposed introduction of channel catfish (Ictalurus punctatus) to New Zealand." Conservation Biology 6, 273-82.
Uvarov, E.B., & Isaacs, A. (1986). The Penguin Dictionary of Science. 6th edition. London, UK: Penguin Books. iii+468pp.
van Wensem, J. (1989). "A terrestrial micro-ecosystem for measuring effects of pollutants on isopod-mediated litter decomposition." Munawar et al., 507-16.
Warner, F. (1992). "Introduction." Royal Society, 1-12.
Willis, K.G., & Garrod, G.D. (1993). "Valuing landscape: a contingent valuation approach." Journal of Environmental Management 37, 1-22.
Zaelke, D., Orbuch, P., & Housman, R.F. (eds..) (1993). Trade and the Environment: Law, Economics and Policy. Washington, DC, USA: Island Press. xv+319pp.
RRojas Research Unit/1997