Application of multivariate statistics to ecotoxicological field studies
Ecotoxicological field studies are now an established component of ecological risk assessments. The purpose of such studies is usually to evaluate, under field conditions, the potential for unacceptable effects on populations or communities within the ecosystem. Recent symposia in Environmental Toxicology and Chemistry have discussed the performance and use of community and ecosystem studies in risk assessment for a variety of ecosystem types [1, 2]. Analyzing and interpreting the results of these studies remains one of the major challenges. It requires not only in-depth ecological knowledge , but also tools for presenting and analyzing large, complex data sets in a meaningful way. Furthermore, since these studies should be focused on effects at the community and ecosystem levels, analysis and interpretation should be performed at the same level.
Analysis of field studies has relied mainly on traditional univariate statistics, which can bring a number of problems relating to power and experimental variability. It is often limited to a single or a handful of variables at one time, making overall assessment of effects on the community or ecosystem difficult. Furthermore, with the vast number of potentially confounding variables that can affect population dynamics, such approaches can lead to problems in determining cause and effect.
Using techniques that deal with a large number of variables at one time allows inferences to be drawn about responses to the toxicant and makes presentation of data in a tractable form possible. Multivariate statistical approaches offer such a solution, and now a range of tools are readily available to most ecotoxicologists. Previously, these may have been beyond the grasp of most ecotoxicologists either because of computational demands or impenetrable statistics and software. However, many of these techniques have their origins in community ecology and are therefore readily applicable to ecotoxicological field studies. Issues and techniques were discussed at a workshop held prior to the 1996 SETAC Annual Meeting in Washington, DC, USA; the papers in this special section represent a portion of the results of this meeting.
Analytical and interpretational issues
One area of considerable effort over recent years has been in aquatic meso- and microcosm studies . These studies have evolved over the last decade from very large and poorly targeted pond experiments to much smaller and more carefully designed trials. However, the number and frequency of measurements often remains large and the number of replicates is sometimes small. This can lead to considerable difficulties in data processing and statistical analysis, especially when using univariate techniques such as analysis of variance. Due consideration should be given to the experimental design and optimum taxonomic level of identification for these studies.
In contrast to experimental studies, field monitoring studies (e.g., before and after chemical exposures, upstream-down-stream measurements) present difficulties related to the sometimes semiquantitative nature of the data and the limited scope for true control sites, replication, and experimental manipulation . The almost inevitable presence of confounding factors that cannot be controlled is another difficulty. Consequently, these restrictions have led to the use of a wider range of data evaluation techniques, including indexes of ecosystem health and some multivariate measures. These approaches tend to have a greater degree of ecological meaning, but interpretation and establishing causality can be difficult. This means that the natural range of variability must be established at reference sites before contaminant-related effects can be correctly determined and interpreted.
Ecotoxicological field studies are sometimes used in the regulation of pesticides and are triggered when laboratory tests have indicated a potential risk . Studies would be improved by a much clearer articulation of management goals and desired endpoints; obviously, field studies are useful only if they genuinely address real policy needs. They also must be capable of detecting changes of a magnitude and rate that are ecologically relevant. Improving the ability to extrapolate from the experimental community to real ecosystems would be another enhancement. A serious difficulty with ecotoxicological field studies is the problem of communicating the often complex results to nonspecialists, who ultimately have to make registration decisions.
A variety of multivariate techniques offer potential solutions to these analytical and interpretational problems . Exploratory methods can be used to determine if a putative pollution in the field is having a measurable impact. Appropriate methods are based on classification and ordination such as clustering, principal components analysis, multidimensional scaling, correspondence analysis, and detrended correspondence analysis. Explanatory methods attempt to relate field effects to measured environmental variables, both natural and anthropogenic. Appropriate methods are based on canonical analysis such as biplots, canonical correspondence analysis, redundancy analysis, and principal response curves. Confirmatory hypothesis testing methods (the multivariate equivalents of ANOVA) may be used to assess the significance of differences between treatments in manipulative field experiments and in laboratory bioassays of field-collected toxicants. Appropriate hypothesis testing methods are multiway analysis of variance, Hotelling's T2, Wilks λ, and permutation tests such as analysis of similarity and Monte Carlo tests.
Analyzing ecotoxicological field studies with multivariate techniques has some clear advantages. Community-level approaches have more ecological relevance than studies at lower levels of biological organization, and as yet, no compelling evidence suggests that they are any less sensitive at detecting the biological effects of pollution, especially when multivariate analyses are applied. Moreover, since multivariate analyses of multispecies data contain a plethora of information, they are more likely to discriminate between treatments than simple univariate summaries of the same data, which may miss some of the more subtle nuances of community changes. Consequently, these approaches may be more helpful in determining the ecological significance of toxicant impacts and may help the evaluator of the study reach conclusions based on ecologically important effects, a fundamental responsibility in field studies.
Cost-effectiveness is clearly important, and costs can be reduced dramatically by considering taxonomic sufficiency (do we really need to painstakingly identify everything to species level?) and a sampling design appropriate for the subsequent statistical analysis. Multivariate techniques are also ideal for handling large amounts of data and endpoints more efficiently.
Past criticisms of multivariate statistics have focused on their relative inaccessibility and the difficulty of interpreting complicated outputs. In recent times, much has been done to counter these criticisms. Such techniques are becoming increasingly practical because of the recent availability of user-friendly software, for example, the principal response curves method  and the routines in the PRIMER software package . Major steps have also been taken to produce outputs readily interpretable by both ecologists and environmental managers or regulators. Multivariate techniques now provide the ecotoxicologist with powerful tools to visualize and present impacts at the community and ecosystem levels.
A number of new techniques under development may offer additional opportunities for ecotoxicologists. These include certain nonstatistical multivariate approaches such as machine learning of rules and neural networks. Future years will undoubtedly see the wider application of such approaches to ecotoxicological field data.
Conclusions and recommendations
A number of recommendations may help to form the framework for future discussions and activities in this subject area. (1) In certain circumstances, ecotoxicological field studies in risk assessment are still needed. They should address specific objectives relating to effects at the community (and potentially ecosystem) level of organization. (2) Ideally, community-level endpoints should enable determination of cause-and-effect relationships and allow for confounding variables. Covarying data (indirect effects) should also be considered carefully. (3) Although multivariate statistical approaches offer potential solutions to these issues, some guidance on the suitable applications of multivariate techniques would be valuable to ecotoxicologists. (4) Application of multivariate statistics may require the modification of ecotoxicological field study design. Due consideration should be given to appropriate levels of replication, the endpoints to be measured, and the taxonomic resolution required. (5) The level of effort applied to experimental design and statistical analysis should equate the overall level of resources expended in the study. Closer cooperation between statisticians and ecotoxicologists would also link data collection and analysis, and result in improved studies. (6) Integration of multivariate techniques into regulatory assessments will improve the value and efficiency of ecotoxicological field studies for ecological risk assessment.
A key responsibility of the ecotoxicologist is to use these techniques to reach ecologically sound conclusions. Any statistical differences or trends should be interpreted in light of their biological relevance.