Clements: “3357_c010” — 2007/11/9 — 12:41 — page 163 — #1 10 Sublethal Effects 10.1 OVERVIEW Although sublethal effects are often more subtle than those inducing direct mortality, they can equally impact overall community structure. (Bridges 1997) The above statement could not have been made in the 1960s and 1970s without requiring immediate qualification. Then instances of acutely lethal exposures to human products and byproducts were blatant and preoccupied most wildlife and aquatic toxicologists. Acutely lethal exposures still occur (e.g., accidental bird poisonings with pesticides (Stansley and Roscoe 1999)) but our attention is being drawn increasingly to effects that do not produce outright death. This is not to imply that sublethal effects were completely ignored in the past: published contributions to the sublethal effects literature began well before the 1960s. Indeed, sublethal effects featured prominently in the opening chapter of Rachel Carson’s Silent Spring (1962). Sublethal effects were simply treated as secondary in importance in the 1960s and 1970s. However, interest in and effort spent on sublethal effects has appropriately comeintobalance duringthe past twodecades with thosefocused on lethaleffects. Most sublethal studies explore one or more of five fitness-related features of individuals: reproduction, growth, development, behavior, and physiology. Very often, growth and reproduction are examined simultaneously. Significance tests are for situations where we do not understand, in any theoretical sense, what is happening. (Hacking 2001) A few key points can be made at the onset about the current sublethal effects literature based on a quick survey of 114 randomly selected studies published between 1968 and 2006. Eighty-two percent of these papers applied experimental designs appropriate for analysis of variance (ANOVA) or analysis of covariance (ANCOVA), and did not draw on available ecological models to quantify or interpret results. Most were designed to detect change in response to increasing exposure level and interpreted results in that context. Basic hypothesis testing dominated analyses though descriptive regression models were common. Theory-based experimental designs were applied in fewer than 10% of the surveyed publications. This is surprising given the observation that 47% of these papers attempted to link results to fitness, demographic, or bioenergetics consequences in their discussions. This observation and the above quote suggest that the pervasive use of significance testing results from a lack of trust in or knowledge of available theory that could be applied to design better experiments. Perhaps one cause of this incongruity is the longstanding regulatory stance that sublethal effects are best addressed with hypothesis testing. Pragmatic coping with past problems gives justification for the emergence of this position but its continued maintenance becomes less justifiable with each passing year. A clear tendency away from the dominance of hypothesis testing of sublethal effects now seems to be emerging in Environmental ProtectingAgency (EPA) and other agency documents. Because the field is currently shifting relative to how we deal with sublethal effects, the reader will find considerable discussion of hypothesis testing in this chapter but will also find very rel- evant theory and studies in the chapters that follow. The theory contained in those chapters bridge 163 © 2008 by Taylor & Francis Group, LLC Clements: “3357_c010” — 2007/11/9 — 12:41 — page 164 — #2 164 Ecotoxicology: A Comprehensive Treatment Available energy Growth Reproduction Time between life cycle events Survival Phenotype- and environment-dependent trade-offs/optimization Conditional maximum fitness Behavior FIGURE 10.1 Sublethal and lethal effects are interlinked and patterns of effects should be interpreted within this context. Strategies evolve within the phylogenetic constraints of an organism in order to optimize Darwinian fitness in a range of environments. Coordinated reallocation occurs of energy resources to maintain the soma (survival), increase the soma (grow), make transitions among life stages, and reproduce. Behavior also can be modified during the expression of phenotype. A strategy designed by natural selection to maximize fitness will be taken given a particular environment and phenotypic plasticity. the conceptual gulf over which sublethal effects information cannot currently pass without taking on considerable—likely unacceptable—uncertainty and ambiguity. In addition to methodological changes taking place about how we deal with sublethal effects, a less obvious evolution is emerging relative to how we go about interpreting the rapidly growing body of sublethal effect data. Is it best to interpret changes based solely on the mechanisms described in earlier chapters or are there “emergent properties” (i.e., higher-order phenomena) that must be considered too? As we will discuss again in Chapter 16, trade-offs and energy allocations are made in complex ways by individuals faced with variable environments (Figure 10.1). Some of the most important involve allocation of energy among growth, somatic maintenance (survival), andreproduc- tion. Other equally important shifts are associated with the timing of life-cycle events and foraging behavior. Ideally, these trade-offs integrate in a manner that maximizes an individual’s fitness within the confines of the particular environment in which the individual finds itself (Figure 10.2). It is unreasonable to assume that evolved life history strategies and trade-offs associated with optim- izing fitness in changing environments do not also manifest within sublethal effect data sets. In many cases, mechanisms and paradigms associated with such strategies might be equally or more relevant than those associated with lower levels of organization. For example, Brown Sullivan and Spence (2003) conclude from a conventional, lower →higher-level interpretative vantage that some sublethal effects of atrazine and nitrate appear inexplicably to be antagonistic while others are syner- gistic. Perhaps these seemingly contradictory effects from the vantage of suborganismal mechanisms could be interpreted successfully using life history strategy theory. Similarly, Heinz and Hoffman noted from their studies of selenium and mercury effects on mallard ducks (Anas platyrhynchos) that “mercury and selenium may be antagonistic to each other for adults and synergistic to young, even in the same experiment.” Such differences are not easily explained based on suborganismal © 2008 by Taylor & Francis Group, LLC Clements: “3357_c010” — 2007/11/9 — 12:41 — page 165 — #3 Sublethal Effects 165 Growth Survival Timing Reproduction Factor A Factor B Behavior FIGURE 10.2 In differing environments, including those containing toxicant stressors, Darwinian fitness is optimized for an individual of a particular species by shifting energy resources allocated to growth, survival, and reproduction, and changing behavior and the timing associated with achieving life history events such as metamorphosis to a sexual adult. The hypothetical factors A and B shown here can be physical (e.g., habitat), chemical (e.g., salinity or chemical toxicant), or biological (e.g., competitors) factors. mechanisms alone but could be explained and further scrutinized with life history strategy theory. Higher-level controls are exerted on many suborganismal processes and predicted by optimality or life history strategy theories. Regardless, is it possible to effectively address these higher-order phenomena in studies of sub- lethal effects? The unambiguous answer is yes. Handy et al. (1999) explored copper effects on rainbow trout (Oncorhynchus mykiss) locomotion and detoxification in terms of trade-off theory. Knops et al. (2001) examined the balance between metabolic costs of resisting toxicant effects versus those of reproduction and growth. Sibly and Calow (1989) and Atchison et al. (1996) write convincingly about linking ecotoxicology to life-cycle theory and optimization theory. Books such as Stearns’ The Evolution of Life Histories (1992) provide ample detail about relevant processes and concepts. Higher-order phenomena can be integrated into ecotoxicity studies if one seeks out the appropriate methods and theories instead of remaining fixed in the conventional mode of inter- pretation that emerged to address the immediate issues confronting ecotoxicologists in the 1960s and 1970s. Complementary interpretative paradigms are now blending in an exciting way in the area of sub- lethal toxicant effects. An emerging consensus from associated studies is that higher-order processes © 2008 by Taylor & Francis Group, LLC Clements: “3357_c010” — 2007/11/9 — 12:41 — page 166 — #4 166 Ecotoxicology: A Comprehensive Treatment can be as important as those suborganismal mechanisms discussed in earlier chapters. Associated higher-level theories also provide an essential way of translating organismal effects into those to populations. The current blending of concepts in the area of sublethal effects leads to difficulty in presenting the associated information here. A conventional approach will be taken that hopefully will not leave the reader with the false impression that all issues relevant to sublethal effects are covered in this chapter. Previous chapters discussed the mechanistic underpinnings of sublethal effects and chapters that follow will discuss important theories explaining higher-order properties that modify suborganismal processes. This short chapter will generally discuss fitness consequences and then conventional ways of quantifying sublethal effects. The reader is urged to explore earlier chapters for detailed discussion of suborganismal mechanisms giving rise to these effects and later chapters for relevant higher-order mechanisms influencing sublethal effects and responses. Concepts essential to understanding sublethal effects will be missed if those materials were overlooked by the reader. Toxicant exposure often completely eliminates the performance of behaviors that are essential to fitness and survival in natural ecosystems, frequently after exposures of lesser magnitude than those causing significant mortality. (Scott and Sloman 2004) 10.2 GENERAL CATEGORIES OF EFFECTS Sublethal stressor effects that dominate the literature can be discussed in four convenient groupings: growth and development, reproduction, behavior, and physiology. From the first such publication to the most recent, implications remain on how these effects diminish Darwinian fitness. Although not often stated as such, this is the context in which they are interpreted in most regulatory processes. The diversity of these studies can be illustrated below with the brief summary of the surveyed research publications mentioned earlier. 10.2.1 DEVELOPMENT AND GROWTH Slightly less than a third of the 114 research publications that we surveyed described stressor effects on growth and roughly the same fraction described effects on development. Many addressed them together. Developmental studies included thoseexamining conventional teratogenic effects, those emphas- izing sexualdevelopment, and those examiningmore subtledevelopmental effects(e.g., changesin an individual’s behavior due to exposure during development). Early examples of the first type are Pech- kurenkov et al. (1966), D’Agostino and Finney (1974), Ward et al. (1981), Bookhout et al. (1984), and Conrad (1988) who noted morphological changes in fish, crustaceans, or molluscan eggs or lar- vae during exposure to a range of chemicals. More recent studies apply similar approaches but tend to address more relevant than conventional test species (e.g., DeWitt et al. 2006) and stressor (e.g., Moreels et al. 2006, Williams et al. 2003, Wollenberger et al. 2005) combinations. Many recent stud- ies such as Carr et al. (2003), Degitz et al. (2000), Fordham et al. (2001), Rowe et al. (1996), Brown Sullivan and Spence (2003), and Tietge et al. (2005) explore the important issue of stressor effects on amphibian development. Afew emphasize developmental stability (e.g., Green and Lochman 2006), which is described in Chapter 16. These more recent publications also have a greater tendency to frame studies in a mechanistic context; for example, endocrine or reproductive system modification by contaminants (e.g., Boudreau et al. 2004). Those publications that focused on reproductive effects are currently dominated by fish studies (e.g., Bortone et al. 1989, Parrott et al. 2003, Penske et al. 2005, Teather et al. 2005, Toft et al. 2004) but developmental effects to important bird (Fry and Toone 1981), insect (Delpuech and Meyet 2003), and molluscan (Bryan and Gibbs 1991) species are also common. Many emphasize either reduced reproductive fitness consequences or imposex © 2008 by Taylor & Francis Group, LLC Clements: “3357_c010” — 2007/11/9 — 12:41 — page 167 — #5 Sublethal Effects 167 (the development of male characteristics such as a penis in females). Still other studies examined nonmorphological effects. As an example, Samson et al. (2001) explored delayed functional effects to zebra fish (Danio rerio) after exposure to methylmercury, including behavioral effects. Weis and Weis (1987) provide a good, but somewhat dated, review of such developmental effects. Studies of toxicant effects on growth were similarly diverse and slightly more plentiful because growth is relatively easy to measure. As a very welcome exception to their rarity in most other areas of organismal ecotoxicology, plant studies were common. Plant growth studies ranged widely, including those involving terrestrial macrophytes (e.g., Baker and Walker 1989, Boutin et al. 1995, 2000, Fletcher et al. 1988, 1996, Wilson 1988), aquatic macrophytes (Lewis and Wang 1999, Lytle and Lytle 2005, Marwood et al. 2003, Stewart et al. 1999), and microscopic plants (Mayer et al. 2000, McGarth et al. 2004). Reports of hormesis (see Chapter 9) appeared in several plant studies (e.g., Calabrese et al. 1987, Stebbing 1982). Growth as affected by toxicants was also commonly studied for terrestrial (e.g., Coeurdassier et al. 2001, Inouye et al. 2006, Spurgeon et al. 1994) and aquatic (e.g., De Schamphelaere and Janssen 2004, Gray et al. 1998, Ingersoll et al. 1998, Knops et al. 2001) invertebrates. Growth studies of vertebrates included such diverse study groups as birds (Bishop et al. 2000, Hanowski et al. 1997, Woodford et al. 1998), fish (Al-Yakoob et al. 1996, Cook et al. 2005, Hansen et al. 2004, Munkittrick and Dixon 1988, Schmidt et al. 2005), and amphibians (e.g., Diana et al. 2000, Relyea 2004, Schöpf Rehage et al. 2002, Wojtaszek et al. 2004, 2005). 10.2.2 REPRODUCTION Thirty percent of the surveyed sublethal effects publications examined reproduction alone or in combination with another effect. Studies of reproductive effects are so common that standard assays have been established such as those for the fathead minnow (Ankley et al. 2001, Bringolf et al. 2004). In contrast to the older publications such as Arnold (1971) or Bodar et al. (1988), the more recent reproductive effects publications tended to focus less on conventional test species such as Daphnia magna and more on species and exposure scenarios relevant to the particular situation of concern. As examples using aquatic crustacea, Wirth et al. (2002) examined grass shrimp (Palaemonetes pugio) chronically exposed to endosulfan, and Cold and Forbes (2004) studied growth of Gammarus pulex experiencing pulsed exposures to esfenvalerate. Additional examples are easily found for terrestrial species including soil-associated species (e.g., Collembola exposed to arsenic) (Crouau and Cazes 2005), and annelids exposed to chemicals from explosives (Dodard et al. 2005) or metals (Kuperman et al. 2006). Birds arecommon subjects of reproductionstudies as a consequenceof historical events involving avian reproduction such as dichlorodichloroethylene (DDE)-linked eggshell thinning (Hickey and Anderson 1968) and current problems such as selenium’s effect on Kesterson Wildlife Refuge (Cali- fornia) waterfowl and wading birds (Ohlendorf 2002). Typical of well-done field studies of avian reproduction are those ofBishop et al. (2000)of the possible impact of apple orchard-associatedpesti- cides on Tree swallows (Tachyecineta bicolor) and Eastern bluebirds (Sialia sialis), and Ohlendorf (2002) who surveyed birds associated with a selenium-contaminated region of San Joaquin Valley. Also typical are field studies of birds particularly prone to chronic exposure such as piscivorous birds exposed to dietary mercury (e.g., Elbert and Anderson (1998) and Meyer et al. (1998)). Similarly, laboratory studies of toxicant effects on avian reproduction are accumulating in the literature (e.g., Heinz and Hoffman (1998)). Standard methods have been established for quantifying avian repro- ductive effects although Mineau et al. (1994) suggest important shortcomings in these tests relative to predicting effects in the field. 10.2.3 BEHAVIOR Surprisingly, nearly 50% of the surveyed publications contained descriptions of behavioral changes of one sort or another. Some were straightforward reports of changes in locomotor behavior. © 2008 by Taylor & Francis Group, LLC Clements: “3357_c010” — 2007/11/9 — 12:41 — page 168 — #6 168 Ecotoxicology: A Comprehensive Treatment Zebrafish (Brachydanio rerio) general swimming (Grillitsch et al. 1999, Vogl et al. 1999), amphipod (Gammarus lawrencianus) swimming direction (Wallace and Estephan 2004), annelid (Lumbriculus variegates) helical swimming (O’Gara et al. 2004), and woodlouse (Oniscus asellus) movement (Bayley et al. 1997) are a few of the locomotion changes related to toxicant exposure in studies. Still other straightforward studies assessed an individual’s ability to simply avoid high concentrations of toxicants (e.g., Kynard 1974, McCloskey et al. 1995, Roast et al. 2001, Sprague 1968). Other behavioral studies focus on endpoints with implicit connection to an individual’s general fitness. Examples include the influence of mercury on the foraging behavior of fish (Weis and Khan 1990) and the sediment reworking activities of a benthic oligochaete (Landrum et al. 2004). Social (intraspecies) interactions are another important set of behaviors that can be changed by toxicant exposure. These include aggression (e.g., Janssens et al. 2003), social structuring (e.g., Sloman et al. 2003), schooling (e.g., Nakayama et al. 2005), and mating (e.g., Hunt and Warner Hunt 1977) behaviors occurring within groups of individuals of the same species. Finally and as detailed in later chapters, interspecies interactions are also important, including balancing activities associated with foraging and predator avoidance (e.g., Hui 2002, Perez and Wallace 2004, Preston et al. 1999, Riddell et al. 2005, Schulz and Dabrowski 2001, Sullivan et al. 1978, Tagatz 1976, Webber and Haines 2003). All of these behaviors influence the overall fitness of an individual with or without the influence of toxicant exposure. 10.2.4 PHYSIOLOGY Physiological effects can decrease fitness directly or indirectly. Even the behavioral changes just described create the potential for physiological shifts. For example, Sloman et al. (2003) noted that subordinate rainbow trout (O. mykiss) accumulated more copper than dominant rainbow trout, notionally creating a difference in potential for physiological effects in different trout within a social hierarchy. Shifts in physiology, including shifts associated with energy expenditure, reduce an individual’s options relative to optimizing energy allocation. Such energetic costs have been measured directly in carp (Cyprinus carpio) exposed to copper (De Boeck et al. 1997); bivalve molluscs (Pisidium amnicum) and salmon (Salmo salar) exposed to pentachlorophenol (Penttinen and Kukkonen 2006); and bivalve molluscs (P. amnicum), Chironomidlarvae(Chironomus riparius), and oligochaetes (L. variegates) exposed to 2,4,5-trichlorophenol (Penttinen et al. 1996). Consequent to all of the issues described above, an increasingly common interpretation of sublethal effects is based on energy allocation (e.g., Handy et al. 1999). 10.3 QUANTIFYING SUBLETHAL EFFECTS Results of almost all life-cycle, partial life-cycle, and early life-stage toxicity tests have been calculated using hypothesis tests in conjunction with analysis of variance to detect statistically significant differences from the control treatment, whereas results of almost all acute tests have been calculated using regression analysis. Because the experimental designs for these two types of toxicity tests usually are very similar, both hypothesis testing and regression analysis can be used to calculate results of both acute and chronic toxicity tests. (Stephan and Rogers 1985) Conventional tests for sublethal effect consist of replicate groups of organisms exposed to a series of concentrations or doses 1 for a specified duration. The treatment levels include one or more types of 1 Although sometimes mistaken as synonyms, it is important to remember that concentration and dose are not the same. Concentration is the mass of the chemical per unit of mass or volume of the relevant media to which the organism is exposed. Dose is an amount administered to or entering an individual such as the amount ingested by an individual. The related term, dosage is simply a body mass normalized dose (e.g., 5 mg/kg of body weight). © 2008 by Taylor & Francis Group, LLC Clements: “3357_c010” — 2007/11/9 — 12:41 — page 169 — #7 Sublethal Effects 169 reference (no toxicant) treatment and a series of increasing concentrations or doses. Some protocols specify that effects should be noted at a few durations, not a single one. Sometimes, a contaminated medium such as an effluent, soil, or sediment is mixed with uncontaminated media to produce the graded series of test exposure treatments. A variety of manuals provide details for conducting such tests andanalyzingresults; forexample, thewell-establishedEPAShort-Term Methods for Estimating the Chronic Toxicity of Effluents and Receiving Waters to Freshwater Organisms (EPA 2002) and the recent OECD Current Approaches in the Statistical Analysis of Ecotoxicology Data: A Guidance to Application (OECD 2006). In these manuals, recommended effect metrics are calculated using either hypothesis testing or point estimation methods. Which is the best approach has been vigorously debated for at least two decades (e.g., Stephan and Rogers 1985, Chapman et al. 1996, Crane and Newman 2000); therefore, the salient points of the debate are summarized below. As described in the above quote from Stephan and Rogers (1985), the same data set can be analyzed with these two methods. However, optimization of experimental design is not the same for both methods. On the basis of well-accepted design principles, optimization of hypothesis testing might involve more replicates per treatment but optimizing point estimation by regression might involve spreading the experimental units among more concentrations and selecting concentrations closer to the level of effect for which estimation is being done (Figure 10.3). Or optimization for regression analysis could involve another distribution of experimental units depending on the model being applied. For example, the best distribution of experimental units for an exponential model would be different from that for the simple model depicted in Figure 10.3. More replicates in the control or reference treatments improve power of many hypothesis tests (see pages 17–31 in Cochran and Cox (1957)), but the best distribution of experimental units to produce good regression estimates is dependent on the applied model and the point being predicted (see pages 86–89 in Draper and Smith (1998)). 10.3.1 HYPOTHESIS TESTING AND POINT ESTIMATION The hypothesis testing approach attempts to identify the highest concentration or dose that has no effect (i.e., either a biological or proof-of-hazard threshold). 2 This renders in practice to statist- ically comparing the level of effect measured in a series of experimental treatments to that of the reference or controltreatment(s). Thisis done with conventionalhypothesis tests that compare means, medians, or other distributional qualities for the experimental units of the treatments (see Newman (1995) for detailed descriptions). Statistical tests identify the lowest concentration that is significantly different from the control or reference treatment(s) (i.e., the lowest observed effect concentration or level (LOEC or LOEL)). The highest concentration that is not statistically significant different from the reference or control treatment is the no observed effect concentration or level (NOEC or NOEL). In common practice, the NOEC and LOEC calculated in a full life-cycle or partial life-cycle test 3 are conditionally used to establish “Safe Concentrations.” Appropriately, results from this hypothesis testing approach are increasingly being judged insufficient to estimate safe concentrations unless they are associated with enough statistical power to detect a relevant change in a key ecotoxicological effect and, equally important, their interpretation incorporates appropriate biological theory. 2 The reader should know that, although commonly done, it is not strictly valid to make a judgment about biological thresholds with the usual hypothesis testing methods. Cautious users of hypothesis testing interpret results in a proof-of- hazard context instead: no hazard is assumed if no evidence exists at a particular dose or concentration level. The evidence is a significant deviation from the null hypothesis. The threshold becomes not a strict biological threshold but what could be referred to as a threshold of toxicological concern (DeWolf et al. 2005). The problem is that it is often treated incorrectly as a proof of safety threshold. The reader is invited to read Hauschke (1997) or OECD (2006) for more details. 3 A life-cycle test is one in which key components of individual’s life cycle are assessed for contaminant adverse effects. For example, reproduction, development, growth, and survival of a test species might be examined in a battery of tests and effect metrics determined for each. Because of the expense associated with a full life-cycle test, partial life cycle tests have been developed that focus only at the notionally “most sensitive” stages of an organism’s life cycle, usually the early stages. © 2008 by Taylor & Francis Group, LLC Clements: “3357_c010” — 2007/11/9 — 12:41 — page 170 — #8 170 Ecotoxicology: A Comprehensive Treatment NOEC LOEC Growth rate Treatment concentration 0 10050 25 12.5 Growth rate 024 6810121416 Growth rate expressed as x % drop relative to the control animals EC x ∗ ∗ FIGURE 10.3 A hypothetical case illustrating hypothesis testing and point estimation, and changes to design necessary to optimize effect metric generation. The top panel shows an experimental design with a control treatment and four concentration treatments (12.5, 25, 50, and 100). Each noncontrol treatment has triplicate cages containing five waterfowl each. The mean growth rate for each cage is used as the effect variable. To optimize power ofthe hypothesistesting, the number ofcontrol replicates (cages)is setat the numberof replicates in each noncontrol treatment times the square root of the number of comparisons or √ 4 × 3 = 2 × 3 = 6. (See Newman (1995) or OECD (2006) for details and references.) The NOEC and LOEC are obtained on the basis of the treatments with mean growth rates significantly different (*) from the mean control growth rate. If the experiment was intended to produce an EC x , the design would be optimized in a slightly different manner (bottom panel) because the emphasis would be on producing a good point estimate instead of optimizing power. More treatment concentrations near the suspected EC x might be chosen with fewer replicates within each treatment. Depending on the selected model, the treatment concentrations might also be spaced to optimize estimation. Point estimation methods begin by assigning a level of effect that is deemed unacceptable and then estimating the corresponding concentration or dose. Several approaches can be used, ranging from fully parametric regression modeling to nonparametric estimation. Models range from simple dose–response to very complex models. They might or might not include thresholds, natural baseline response levels, or hormesis. The OECD (2006) report mentioned above has an especially good discussion of such models although discussion of biologically-based models is restricted to only one of several potential mechanistic models. Point estimation methods predict the concentration or dose associated with a given level of effect. Often, this involves prediction from a regression model fit to the ecotoxicity data set. Predictions are © 2008 by Taylor & Francis Group, LLC Clements: “3357_c010” — 2007/11/9 — 12:41 — page 171 — #9 Sublethal Effects 171 referred to in such terms as the EC x , the effective concentration calculated to have an x% change in the response. 4 Models fit to effects data sets are often those described in Chapter 9. As already mentioned, the virtues and shortcomings of these two approaches have been and still are debated. This debate extends back to the origins of these approaches, that is, back to human health assessment where No Observed and Lowest Observed Adverse Effect Concentrations (NOEAC and LOEAC), and Benchmark Doses are still used in a similar fashion. There are situations in which point estimation is a compromised tool. The first major drawback of point estimation is highlighted in the above Hacking quote. Hypothesis testing is preferable to modeling if one has no understanding of the relationship between the effect and the toxicant concentration (or dose). Second, hypothesis testing might be the only option if the variability in the data set does not allow one to identify and fit a model. Several ecotoxicologists have argued that a third shortcoming is that an x must be defined a priori with point estimation and insufficient insight often exists with which this has to be done with acceptable certainty. However, the hypothesis testing approach does not avoid the crucial issue of determining what is a biologically unacceptable level of effect except during its misapplication (i.e., when statistical significance is mistakenly equated with biological relevance). Hypothesis testing simply puts the question off. Is it not better to confront these uncertainties at the onset of an investigation? Fourth, some of the model fitting techniques are necessarily iterative and situations exist in which they fail to converge on an acceptable solution. The associated parameter estimates and predictions are unacceptable in that case and the hypothesis testing might be the only tool available. The shortcomings of hypothesistesting are more commonly discussed than thosefor point estima- tion. Point estimation is oftenpresented as a superiortechniquethat will eventually replacehypothesis testing as the preferred meansofproducing sublethal effects metrics. Most of the argumentstoabolish the hypothesis test-associated metrics (e.g., Crane and Newman 2000, Kooijman 1996, Laskowski 1995) emerge from the pervasive misapplication of hypothesis tests and misinterpretation of test res- ults, not any fundamental disagreement with Neyman–Pearson theory. The first shortcoming is that statistical significance is not a reliable indicator of biological significance or relevance of an effect. The ecotoxicological literature is replete with instances in which the two are confused. Newman and Unger (2003) refer to this pervasive confusion as the maulstick incongruity. As an example, the estimation of a hazardous concentration (HC p ) for p percent of species in a community from a collection of NOEC values of those species (e.g., Van Straalen and Denneman 1989) requires one to assume that the NOEC is the concentration at or below which there is no biologically signi- ficant effect. This is clearly an overextension of the concept because the NOEC/LOEC values are extremely dependent on the experimental design, variation in the data, and the particular significance test applied. The subtle extension of the NOEC/LOEC values to vaguely infer hazard or risk can be found in many sources such as the following quote: The parameter p in HC p was considered equivalent to risks estimated for industrial accidents, cancer risks from radiation, etc. Consequently, HC 5 could be considered as a concentration with an ecological risk of 5%, with risk in this case being the probability of finding a species exposed to a concentration higher than its NOEC (Van Straalen and van Leeuwen 2002) Although such statements are motivated by well-intended pragmatism, it is difficult to separ- ate inferences of statistical significance and biological relevance with such effect metrics. Does exceedance of an NOEC constitute an ecological risk as inferred? Kooijman (1987) and Newman (1995) articulated concern about making such inferences in hazardous concentration estimations. 4 The EC x is similar to the human toxicologist’s Benchmark Dose (BMD) approach that uses regression model predictions for a specified effect level (benchmark response) instead of hypothesis testing. Often the lower 95% confidence limit for such an estimated BMD (BMDL) is used to set exposure limits in human risk assessment (Crump 1984, Falk Filipsson et al. 2003). © 2008 by Taylor & Francis Group, LLC Clements: “3357_c010” — 2007/11/9 — 12:41 — page 172 — #10 172 Ecotoxicology: A Comprehensive Treatment Further, as a second shortcoming, the NOEC/LOEC metrics are relatively static metrics whereas a regression model can estimate a range of effect concentrations as our knowledge of the level having a relevant biological effect changes with time. Several related concerns have been added to the two already mentioned. Third, the values of the NOEC/LOEC metrics are dependent on design and statistical features of the process, not simply the biological properties being studied. A fourth criticism is that the values of the NOEC/LOEC can change depending on the power of the specific hypothesis test applied to the data set. A related fifth criticism is that the conventional methods used to estimate these metrics generally have the power to detect effects at the level of roughly a 20% change, yet some effects will have biological relevance with much smaller changes. However, this criticism could be addressed to a degree by changing the conventional design. The sixth and seventh criticisms are related to the metrics derived from the hypothesis tests. The manner in which the NOEC/LOEC metrics are derived from the results dictates that they can only take on the value of a treatment. The spacing and selection of the treatments have a strong influence on the NOEC/LOEC values (criticism 6). Next, because they can only take on the value of an experimental treatment, a standard error cannot be produced for the metric estimate (criticism 7). The last several criticisms combine in practice in such an undesirable way that poor design and/or wide within-treatment variability are rewarded with higher NOEC/LOEC metrics (criticism 8). Relative to the conservative application of the effect metrics during environmental decision making, it would be preferable if the opposite were true. The ninth criticism is that the conventional value of .05, or perhaps .01, for the Type I error (α) is an arbitrary one. A critical biological effect with an associated p of .06 might be ignored while a trivial biological effect with a p of .01 might be used to generate the NOEC/LOEC metrics. Obviously, this last criticism is invalid if proper biological insight and judgment were integrated into the procedure including appropriate changes to error rates. Unfortunately, application of such insight is only now becoming obligatory in reports and publications. Statisticians are often stunned by the over-zealous use of some particular statistical tool or methodology on the part of an experimenter, and we offer the following caveat. Experimenters, when you are doing “statistics” do not forget what you know about your subject-matter field! Statistical techniques are most effective when combined with appropriate subject-matter knowledge. The methods are an important adjunct to, not a replacement for, the natural skill of the experimenter. (Box et al. 1978) A tenth criticism relates to treatment assignment and associated error. As generally described by Montgomery (1997), the levels of treatment can be inaccurate in many experiments (e.g., all nominal 100 mg/L treatment replicates are not actually 100 mg/L when measured such that non- trivial differences occur in “replicate” concentrations) and regression methods are more applicable in such cases. An eleventh criticism from Stephan and Rogers (1985) relates to the manner in which sublethal effects testing is done. Often the same experiment is used to test for significant effects of growth, reproduction, and other sublethal effects. The question should be answered in such a case of whether or not the separate hypothesis tests for each effect are independent. A very strong argument could be made that the experimentwise Type I error rate should be adjusted because the tests are not independent. Stephan and Rogers (1985) suggest a twelfth shortcoming of the hypothesis testing approach. The models used for hypothesis tests are rudimentary ones that do not provide ecotox- icologists with an avenue for extending explorations to other models more directly relevant to the biological mechanism or specific context for which inferences are to be made. This criticism will likely become less serious as ecotoxicologists slowly come to appropriately balance the use of hypo- thesis testing and point estimation from biologically well-founded models. Hypothesis testing would then be an invaluable first step in a progression of studies, ending in ecotoxicologically meaningful point estimates. A quick review of the materials just presented results in many more shortcomings for the hypo- thesis test approach than for the point estimation approach. This does not mean that point estimation © 2008 by Taylor & Francis Group, LLC [...]... Bonferroni adjustment of Type I error rates Quantal data Transform? Step-down Comparison to control Parametric Nonparametric Parametric Nonparametric Poisson trend Williams Bartholomew Jonkheere-Terpstra Cochran-Armitage Mantel-Haenszel Dunnett’s Fisher’s exact test with Bonferroni-Holm (BH) adjustment Steel’s many-one with BH adjustment Wilcoxon rank sum with BH adjustment FIGURE 10. 5 Alternative flow chart... tests, contributing to a large data base that is not easily linked to theory-based, predictive models 10. 4.1 SUMMARY OF FOUNDATION CONCEPTS AND PARADIGMS • Most studies of sublethal effects apply experimental designs appropriate for ANOVA or ANCOVA and do not take full advantage of available ecological models to quantify or interpret results • Higher-order phenomena can be integrated into ecotoxicity... studies, J Natl Cancer Inst., 96, 434–442, 2004 Wallace, W.G and Estephan, A. , Differential susceptibility of horizontal and vertical swimming activity to cadmium exposure in a gammaridean amphipod (Gammarus lawrencianus), Aquat Toxicol., 69, 289–297, 2004 Ward, G.S., Parrish, P.R., and Rigby, R .A. , Early life stage toxicity tests with a saltwater fish: Effects of eight chemicals on survival, growth, and development... A. W., Changes in the predator-prey behavior of fathead minnows (Pimephales promelas) and largemouth bass (Micropterus salmoides) caused by cadmium, J Fish Res Board Can., 35, 446–451, 1978 Tagatz, M.E., Effect of Mirex on predator-prey interaction in an experimental estuarine ecosystem, Trans Am Fish Soc., 4, 546–549, 1976 Teather, K., Jardine, C., and Gormley, K., Behavioural and sex ratio modification... independent within a concentration/dose treatment For example, catching batches of ten fish from a holding tank and placing each batch into replicate tanks starting at the control tanks and ending at the highest concentration tanks could easily insert an extraneous factor into the test Perhaps the most easily caught fish were the smallest The fish in the experimental tanks would tend to have the smallest fish... OECD (2006).) Not all recommended tests are shown are nonparametric These differences are generally the same for the continuous data but the Tamhane–Dunnette test is recommended if the data are normal yet variances are not equivalent among treatments 10. 3.1.1 Basic Concepts and Assumptions of Hypothesis Tests Several EPAdocuments establish convention for assessing sublethal effects data sets using hypothesis... Statistical Power Analysis for the Behavioral Sciences, 2nd ed., Lawrence Erlbaum Assoc., Hillsdale, NJ, 1988 Cold, A and Forbes, V.E., Consequences of a short pulse of pesticide exposure for survival and reproduction of Gammarus pulex, Aquat Toxicol., 67, 287–299, 2004 Conrad, G.W., Heavy metal effects on cellular shape changes, cleavage, and larval development of the marine gastropod mollusk, (Ilyanassa... some cases as illustrated with the EPA linear interpolation method (Appendix M of EPA (2002)) With this approach, a concentration associated with a specified level of effect is estimated by linear interpolation between treatments Bootstrap confidence intervals are then generated with the treatment replicate data The result is an estimated ICp (inhibition concentration associated with a specified percent... Toft, G., Baatrup, E., and Guillette, L.J., Jr., Altered social behavior and sexual characteristics in mosquitofish (Gambusia holbrooki) living downstream of a paper mill, Aquat Toxicol., 70, 213–222, 2004 Ury, H.K., A comparison of four procedures for multiple comparisons among means (pairwise contrasts) for arbitrary sample sizes, Technometrics, 18, 89–97, 1976 Van Straalen, N.M and Denneman, C .A. J., Ecotoxicological... control treatment than in the noncontrol treatment Also, some deviations from monotonicity in the data can be accommodated with Williams’ test Finally, these documents often do not highlight the enhanced power available if one can use a one-sided instead of a two-sided test (Cohen 1988) The first step in analyzing sublethal effect data with the EPA scheme (Figure 10. 4) is deciding whether the data should . trend Williams Bartholomew Nonparametric Jonkheere-Terpstra Cochran-Armitage Mantel-Haenszel Parametric Nonparametric Fisher’s exact test with Bonferroni-Holm (BH) adjustment Steel’s many-one with . Kosian, P .A. , Makynen, E .A. , Jensen, K.M., and Ankley, G.T., Stage- and species-specific devel- opmental toxicity of all-trans retinoic acid in four native North American Ranids and Xenopus laevis. Toxicol page 166 — #4 166 Ecotoxicology: A Comprehensive Treatment can be as important as those suborganismal mechanisms discussed in earlier chapters. Associated higher-level theories also provide an