Standard laboratory toxicity tests have been used for the last 60 years to predict ecological consequences of chemical exposure. These tests proved extremely useful because of their reproducibility, cost effectiveness, and ability to link exposure to a chemical with effects (e.g., mortality, immobilization, behavior, and reproduction) on individuals of a species . Forbes and Calow , however, concluded that population growth rate is an ecologically more relevant measure of responses to toxicants than the individual-level effects evaluated by most standard toxicity tests. This limitation to extrapolate the results of standard toxicity tests to the population and community levels and hence to the field led to the implementation of semifield testing. Statistically sound semifield testing started in the 1970s with the introduction of replicated regression designs . Aquatic semifield tests were increasingly used in the 1980s, but many of these studies were conducted in relatively large artificial ponds (mesocosms) and thus required large amounts of time and resources to run and were difficult to interpret because of the high biological variability among replicates, probably resulting from their size and the problems associated with the sampling of large systems. After these challenges, the use of semifield tests in the risk assessment of pesticides was extensively debated. Mesocosm studies were no longer required by the U.S. Environmental Protection Agenct (U.S. EPA) following the new paradigm of 1992, whereas in Europe the focus switched to microcosms, smaller systems in which variability among replicates was more easily managed. Much work was also put into improving microcosm design and statistical evaluation (see Van den Brink  for a list of workshop and guidance documents).
One of the main constraints to the interpretation of semifield experiments is the large data sets they yield, covering many species, sampled in many systems, on many sampling dates. Attempts have been made to summarize the species responses observed using indices that described the effects of chemicals on ecosystem health, but these attempts more or less failed to address the regulatory question of whether there were long-lasting effects on populations and communities. Suter , therefore, concluded that it would be better to assess the real array of ecosystem responses so that causes can be diagnosed, future states can be predicted, and effects of treatments can be compared. This type of assessment, however, required the development of multivariate statistical methods for the analysis of the large data sets. To assess the state of the art, a workshop was held in conjunction with the SETAC North America Meeting in 1996 on the application of multivariate statistics to ecotoxicological field studies . One of the methods presented during this workshop was the newly developed principal response curves (PRC) method, which was designed especially for the analysis of results of semifield experiments .
The PRC method is able to separate the three sources of variation in the community composition of samples from semifield experiment (changes in time, differences between replicates, and effects of treatments) and to show differences in species composition (i.e., community structure) resulting from treatment with respect to the control throughout the duration of the experiment. The result is an easy-to-interpret diagram showing sampling date on the x axis and deviations in species composition of the various communities between the treatments and the control (the PRC) on the y axis. After this first PRC diagram, a second one can be extracted that shows the most important deviations in community response from the one shown by the first PRC, present in the data set. The method also allows an interpretation back to the species level and is able to group species based on their response patterns . By doing so, PRC is able to reveal clear treatment-related response patterns among what would otherwise be noisy ecological data. Given that multiple response patterns can be represented by evaluating, when significant, the first two PRC diagrams, the method is also able to represent all or most of the species that respond to the treatment. To be sure that all responses present in a data set are captured, it is recommended that, in addition to being analyzed by PRC, community data are also evaluated by univariate statistics. Because univariate statistics analyze one taxon at a time, whereas the PRC method analyses the whole community data set, it is expected that PRC will be better able to detect significance of effects compared with univariate methods. The power of semifield experiments to detect significant community-level effects is, however, a subject that requires further discussion and investigation. The PRC method can also be applied to parts of a data set, for example, to show the community response of univoltine versus multivoltine insect species. The method has since been widely applied to analyze the results of aquatic semifield experiments and was recommended by several guidance documents for regulatory evaluation of semifield experiments [9, 10]. However, it can be and has been also applied to other community-level data sets, for example, terrestrial field tests, monitoring studies, and even genetic data.
Thus, whereas single species tests failed to describe effects at the population or community level, semifield experiments are able to address effects at these levels of biological organization by providing realism in chemical exposure, habitat presence, food-web processes, and physicochemical composition of the environment. The observed recovery patterns of affected populations, however, sometimes require extrapolation to scenarios with different functional connectivity . On the one hand, semifield systems may underestimate the recovery potential of affected obligate aquatic species because of their limited physical size and absence of direct connectivity to unstressed systems. On the other hand, they might overestimate the recovery potential of affected species with an aerial life stage because of the presence of untreated controls from which immigration may occur. In extrapolating the results of semifield experiments to different types of landscape scenarios, ecological models offer the potential for such insights .
Although ecological models have already been used for some time in North America for the risk assessment of pesticides , in Europe only during the last few years major initiatives have been undertaken to improve the use of ecological models for the risk assessment of pesticides [14, 15]. Aside from the extrapolation of recovery processes, Hommen et al.  distinguished four other areas for the application of ecological models in chemical risk assessment: (1) extrapolation of organism-level effects to the population level, (2) extrapolation of effects between different exposure profiles, (3) analysis and prediction of indirect effects, and (4) prediction of bioaccumulation within food chains. Initially, the models focused on the assessment of effects of pesticides on life history parameters and how this affects population growth rate, because this results in more relevant assessments of the impacts of pesticides. In addition, measures of population growth rate combine lethal and sublethal effects, which results of standard toxicity tests cannot do . With the increasing consideration of spatial and temporal variability of exposure, effects and recovery patterns are becoming more important for the risk assessment of pesticides. Individual-based models (IBMs) are popular for assessing the effects of spatial and temporal exposure variability, because they are able to keep track of individuals' intraspecific and interspecific interactions as well as their interactions with the specific landscape and, thereby, are able to evaluate individually specific dynamic exposure patterns . An advantage of IBMs is that they can easily be combined with toxicokinetic−toxicodynamic models and dynamic energy budget models, which are both able to provide a mechanistic linkage between exposure and effects when exposure patterns are highly variable in time (e.g., Ashauer ). A great challenge for the future is to implement ecological models into the risk assessment framework by building confidence among users and evaluators of such models.
Although the use of semifield experiments in the ecological risk assessment of chemicals is still sometimes debated, semifield experiments constitute the only experimental means to study population- and community-level effects of pesticides in a realistic environment. Semifield tests also provide data above the level of the population, for example, indirect effects resulting from interactions between species. The PRC method serves as an excellent statistical tool with which to analyze and display effects on community structure. However, in combination with semifield tests, mechanistic community and food-web models would allow a better understanding of these complex studies and extrapolation of the results to nontested environmental scenarios. By combining ecological experimental and modeling approaches, a potentially robust set of higher tier tools will be available for assessing the effects of spatially and temporally varying exposure patterns and subsequent recovery.