Large-Scale Management Experiments and Learning by Doing Author(s): Carl J. Walters and C. S. Holling Source: Ecology, Vol. 71, No. 6 (Dec., 1990), pp. 2060-2068 Published by: Wiley on behalf of the Ecological Society of America Stable URL: https://www.jstor.org/stable/1938620 Accessed: 08-01-2019 16:56 UTC REFERENCES Linked references are available on JSTOR for this article: https://www.jstor.org/stable/1938620?seq=1&cid=pdf-reference#references_tab_contents You may need to log in to JSTOR to access the linked references. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org. Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at https://about.jstor.org/terms Wiley, Ecological Society of America are collaborating with JSTOR to digitize, preserve and extend access to Ecology This content downloaded from 129.219.247.33 on Tue, 08 Jan 2019 16:56:53 UTC All use subject to https://about.jstor.org/terms 2060 SPECIAL FEATURE Ecology, Vol. 71, No. 6 Ecology, 71(6), 1990, pp. 2060-2068 ? 1990 by the Ecological Society of America LARGE-SCALE MANAGEMENT EXPERIMENTS AND LEARNING BY DOING' CARL J. WALTERS Resource Ecology, University of British Columbia, Vancouver, British Columbia, Canada V6T I W5 C. S. HOLLING Department of Zoology, University of Florida, Gainesville, Florida 32611 USA Abstract. Even unmanaged ecosystems are characterized by combinations of stability and instability and by unexpected shifts in behavior from both internal and external causes. That is even more true of ecosystems managed for the production of food or fiber. Data are sparse, knowledge of processes limited, and the act of management changes the system being managed. Surprise and change is inevitable. Here we review methods to develop, screen, and evaluate alternatives in a process where management itself becomes partner with the science by designing probes that produce updated understanding as well as economic product. INTRODUCTION Most of the world's ecosystems are affected to some degree by harvesting and related activities aimed at particular "resource" types or species. But in no place can we claim to predict with certainty either the ecological effects of the activities, or the efficacy of most measures aimed at regulating or enhancing them. Every major change in harvesting rates and management policies is in fact a perturbation experiment with highly uncertain outcome, no matter how skillful the management agency is in marshalling evidence and arguments in support of the change. Practicing resource managers have long been aware of this point, and have tried to invest in adequate monitoring and evaluation programs even while maintaining a public stance of confidence in their predictions. Recently, it has become fashionable to admit at least some degree of ignorance and to label substantial management initiatives as experiments, even when a scientist would shudder at how poorly they are designed. There are major opportunities for research ecologists to become involved in the design and conduct of such experiments, to the mutual benefit of scientists, managers, and resource users. This paper discusses some challenges for justifying and designing experimental management programs. The first challenge is to demonstrate that a substantial, deliberate change in policy should even be considered, given the alternative of pretending certainty and wait- ing for nature to expose any gaps in understanding. A second challenge is to expose uncertainties (in the form of alternative working hypotheses) and management decision choices in a format that will promote both intelligent choice and a search for imaginative and safe experimental options, by using tools of statistical decision analysis. A third challenge is to identify experimental designs that distinguish clearly between localized and large-scale effects, and hence, make the best possible use of opportunities for replication and comparison. A fourth challenge is to develop designs that will permit unambiguous assessment of transient responses to policy changes, in the face of uncontrolled environmental factors that may affect treated and reference experimental units differently. Finally, there is need for imaginative ways to set priorities for investments in research, management, and monitoring, and for design of institutional arrangements that will be in place for long enough to measure large-scale responses that may take several decades to unfold. PASSIVE VS. ACTIVE ADAPTATION There are three ways to structure management as an adaptive process (Walters 1986): (1) evolutionary or "trial and error," in which early choices are essentially haphazard, while later choices are made from a subset that gives better results; (2) passive adaptive, where historical data available at each time are used to construct a single best estimate or model for response, and the decision choice is based on assuming this model is ' For reprints of this Special Feature, see footnote 1, page 2037. correct; or (3) active adaptive, where data available at each time are used to structure a range of alternative This content downloaded from 129.219.247.33 on Tue, 08 Jan 2019 16:56:53 UTC All use subject to https://about.jstor.org/terms December 1990 LARGE-SCALE PERTURBATIONS 2061 response models, and a policy choice is made that re- step in an "Iterative Testing Process" to reshape hy- flects some computed balance between expected short- drological regimes over time so as to permit recovery term performance and long-term value of knowing of as much of the natural Everglades ecosystem as pos- which alternative model (if any) is correct. Most the- sible (S. S. Light, J. R. Wodraska, and S. Joe, unpub- oretical literature on resource management is aimed at lished manuscript). providing single best predictions of policy choice, and Would the Everglades Rainfall Plan have been hence, presupposes that a passive strategy is best. Per- adopted under an active adaptive planning process in- turbation experiments are most likely to arise when an volving deliberate development of a range of alterna- active adaptive strategy is adopted; the main use of tive hypotheses about wading bird dynamics and a active strategies has been in agriculture (field tests, ro- corresponding evaluation of experimental policy op- tation policies) and fisheries (varying harvest rates, tions? At least two broad alternatives deserve consideration. First, at least one species (White Ibis, Eudoc- hatchery systems). There are two fundamental objections to passive imus albus) may have shifted its nesting away from the adaptive policies. First, they are likely to confound management and environmental effects. For example, (Frederick and Collopy 1988); perhaps changes in such Everglades to more northern areas in the Carolinas in fisheries with long monitoring histories (50 + yr), we areas, rather than deterioration in the Everglades, are still have bitter debates about the relative importance partly responsible for the decline. Second, Everglades of fishing and environmental factors in driving popu- nesting colonies were historically concentrated in coastal lation declines and cycles (Walters and Collie 1988). mangrove areas, from which the birds could initially Second, passive policies may fail to detect opportu- move out to forage in "short hydroperiod" marshes nities for improving system performance if the "right" around the system margin and later use the existing model and the "wrong" model predict the same re- marsh core and estuary as seasonal drying progressed. sponse pattern when the system is managed as though Now birds nest mainly in interior water storage (con- the wrong model were correct. Passive adaptive approaches sometimes do result in servation) areas. The margin areas are mostly outside the area impacted by the Rainfall Plan, and have been informative experiments. For example, much of the much affected by drainage for agricultural and urban Florida Everglades system has been lost to agricultural development. They are not currently used consistently development, and water regimes in the remaining marsh by the birds, but they may have been critical for nesting have been altered by a system of upstream dikes and success in the past. If either the "other opportunities" canals developed to permit storage, diversion, and flood or "marginal areas" hypothesis is correct, populations control. These changes had many effects on aquatic will fail to recover under the Rainfall Plan and the communities and vegetation patterns, but the most monitoring program in the experimental area will give publicized one has been a drastic decline in wading no clues as to why the plan failed. An actively adaptive bird populations (Ogden 1978, Kushlan and Frohring 1986, Frederick and Collopy 1988). Efforts to provide mental and monitoring program on a substantially dif- more water to the Everglades National Park, through a quota delivery plan in the 1970s, did not reverse the decline. By the early 1980s, a consensus had emerged that the basic causes of the decline were overall habitat loss and reshaping of the seasonal hydropattern in ways planning process might have resulted in an experi- ferent spatial scale, and a qualitatively different set of manipulations, than is presently considered necessary or feasible. EMBRACING UNCERTAINTY: ASSESSING THE VALUE OF EXPERIMENTAL DECISION CHOICES that caused nesting failures. In 1983, Congress passed the Fascell Bill authorizing tests of alternative water The design and justification of experimental man- delivery plans to the park for 2 yr, and this bill has been extended to the present (1989). In 1985, an "ex- agement plans involve a more complex assessment of risks and benefits than scientists might consider in de- perimental" water release plan, called "The Rainfall Plan," was adopted. Under this plan, water is dis- veloping plans for obtaining statistically significant re- charged to the Park so as to restore as "natural" a to involve a more elaborate and productive interplay sults. In particular, the process of policy design needs between ecologists and decision makers than has trapattern of seasonal and interannual variation as posditionally been seen (Holling 1978). This section illussible. This plan is passively adaptive in the sense that trates some key issues and complexities with an exit was justified by assuming that a single best hypothesis ample from fish and wildlife harvest policy design. (animals require a natural pattern) is correct, yet is also a disturbance experiment that will cause a substantial change in the timing and distribution of water flows into the Park relative to the past two decades (Kushlan For many harvested populations, managers have been "successful" in establishing regulatory programs (re- strictions, monitoring, enforcement) to provide rela1987). It is hoped that the plan will represent the first tively stable population sizes and harvests. Usually it This content downloaded from 129.219.247.33 on Tue, 08 Jan 2019 16:56:53 UTC All use subject to https://about.jstor.org/terms 2062 SPECIAL FEATURE Ecology, Vol. 71, No. 6 is claimed that the current population size is in some malized the values so that the status quo model/policy sense "optimum," such that lower or higher sizes would has a value of 1.0. A rather tedious modelling process on average be less productive of a harvestable excess is usually involved in obtaining credible estimates. A each year. This claim is often supported by estimation technically difficult aspect of such simulations is in of population model parameters from historical data. representing future learning under the experimental Now suppose independent analysis based on "hab- policy option(s); it is necessary to simulate the gath- itat capacity" or alternative parameter estimation pro- ering of noisy data, updating over time of odds placed cedures suggests that the optimum population size has on the alternative hypotheses (i.e., Bayesian learning in fact been grossly underestimated, so the current pop- process), and the eventual shift to an optimum policy ulation size is far below optimum. For example, An- when the correct model becomes reasonably certain. derson (1975) presented production parameter esti- For further details on the modelling methods involved mates for the Mallard (Anas platyrhynchos) suggesting here, see Walters (1986). that the continental population should perhaps be dou- The key policy issue is with the possible outcomes ble the level targeted by management in the 1970s. A of an experimental policy. One possibility is a loss in series of analyses of the Fraser River sockeye salmon short-term yield with no compensating long-term in- (Oncorhychus nerka) have suggested that spawning crease (0.5), and the other is to have the short-term stocks should be far larger than at present (Collie and loss more than balanced by long-term gains (2.0). Of Walters 1987), and that the present value of moving course, if we had calculated the values while using a to a higher stock level could be as much as a half- high discount rate for future harvests (short planning billion dollars to the fishing industry. horizon, no interest in long-term resource husbandry), The usual approach to these cases has been to engage the second column would also have been small. Man- in a "battle of the models," with some scientists and agement experimentation is often meaningless in set- managers defending the status quo and others calling tings where no value is placed on the long-term utility for harvest reductions to allow populations to rebuild of experimental results. to more productive levels. In such battles, the status How should decision makers react to these values? quo argument usually prevails, since it is a "proven" One possibility is to simply average them using some strategy (low risk) and does not involve short-term pain prior odds placed on the alternative hypotheses; in the to the harvesters. However, there is no way to prove example table, such "expected value maximization" that a higher stock would be more productive without would strongly favor the experimental policy option, actually trying it; no behavioral measurements on the unless very high prior odds were placed on the status current population, or highly localized experiments with quo hypothesis. Another possibility is to look mainly increased density, could be guaranteed to anticipate at the worst possible outcome (0.5) and to place much the full variety of responses (such as colonization of weight on this risk. Such risk averse behavior may be marginal habitats) that might accompany the popula- personally favored by many decision makers, but is tion increase. not easily justified in the context of public policy mak- However, in a few cases, such as the Fraser sockeye, ing where the decision maker is supposed to be rep- an actively adaptive approach to management has been resenting the varied interests of many actors. Decision adopted. The essential step in these cases has been to theorists agree that it is important to identify possible avoid a battle of the models, with key parties (scientists, choices and outcomes (the decision table) objectively, managers) agreeing instead to "embrace uncertainty" and to assess the odds of alternative outcomes (assign by laying out a decision table of possible outcomes probabilities to alternative hypotheses), but they have under different alternative hypotheses and policy not yet produced a generally accepted procedure for choices. The simplest possible table might look as fol- combining the information to arrive at optimum public lows: policy decisions. For further discussion of issues and Alternative hypotheses Current stock Optimum is Policy options is optimum at higher stock Maintain status quo 1.0 1.0 Experimentally increase stock 0.5 2.0 methods, see Raiffa (1968), DeGroot (1970), Keeney and Raiffa (1976), and Lindley (1985). A key value of the decision table is to create frustration over a set of policy choices that all appear bad. This frustration can lead to several very useful reac- tions by both scientists and managers: (1) critical evaluation of the alternative hypotheses, with a more care- Each number in this table is an estimate of the ex- ful eye to identifying others and a serious effort to place pected long-term value of following the policy option better odds on the possibilities through more careful in the row, given that the hypothesis in the column is analysis of historical information; (2) a more precise correct. To simplify the comparison of choices, we nor- appraisal of the value of information associated with This content downloaded from 129.219.247.33 on Tue, 08 Jan 2019 16:56:53 UTC All use subject to https://about.jstor.org/terms December 1990 LARGE-SCALE knowing which hypothesis is correct; (3) a search for PERTURBATIONS 2063 the experimental policy will produce a 20% improve- objective ways to weigh the risks and benefits of ex- ment per unit per year. This table would not favor perimentation; and, perhaps most important from a experimentation if all units must be treated alike be- scientific point of view, (4) a search for imaginative cause the baseline policy has higher expected value policy options that might permit the uncertainties to than the experimental policy unless a probability of be resolved through more focused and perhaps less 2/7 or less is assigned to the conservative hypothesis. risky experimentation. There is much for scientists and Suppose the management agency can subject n units decision makers to learn from one another by working each year to the experimental policy (0 < n < N). For together on all of these steps. what n will the expected long-term value from all N For simple decision-making situations involving units be maximum? It may appear that this question harvest management, there have been some attempts can be answered without reference to time, since there to compute optimum feedback policies while explicitly are assumed to be no persistent effects of treatment. accounting for the effect of informative disturbances However, there is a very important temporal dynamic on future management performance (Mangel 1985, that links decision making from year to year, the "in- Walters 1986). These attempts examine the interplay formation state" measure p(t) that summarizes the odds over time of population dynamics and Bayesian learn- placed on the conservative hypothesis by the decision ing, using the methods of stochastic dynamic program- maker in year t. p(t) will not change over time if n = ming to find an optimum policy (Walters 1981, Mangel 0, but will move toward 0.0 or 1.0 if n > 0, depending and Clark 1988). Computational experience to date on which hypothesis is correct. The effect of n > 0 in indicates that the best policy choice for any year will any year is thus to leave the decision maker at time t be either to ignore uncertainty (passive adaptive) or to + 1 with different (on average lower) odds of making make a fairly dramatic and informative experimental the wrong decision in year t + 1. The dynamics of p(t) disturbance; minor experiments are not favored be- are given by Bayes theorem as p(t + 1) = L(t)p(t)/P(t), cause they erode average performance without signif- where L(t) is the likelihood of the responses measured icantly improving learning rates. in year t given the conservative hypothesis and P(t) is CHOOSING THE OPTIMUM NUMBER OF EXPERIMENTAL REPLICATES The balance of learning and risks often does not favor the total probability of the responses: P(t) = L(t)p(t) + L'(t)[l - p(t)], where L'(t) is the likelihood of the data given the optimistic hypothesis. The details of the L(t) calculation need not concern us here. A key point experimental disturbances in single, unique, managed is that some assumption about the likelihood of dif- systems. However, this conclusion changes drastically ferent observed outcomes is necessary for predicting when there is a collection of similar units (lakes, dis- the behavior of p(t), and hence, assessing an optimum tinct populations, areas) that can be managed inde- value of n. pendently. A key question in such situations is how Given the above decision table and a likelihood large an experiment to conduct. Some methodological function L(t), the optimum value of n can be computed issues involved in answering this question can be il- as a function of p(t) by the methods of stochastic dy- lustrated with a simple hypothetical example. Suppose there are N units that are all being managed namic programming (Walters 1981, 1986, 1990, Mangel and Clark 1988). The concept behind the compu- with the same baseline policy, and someone identifies tation is to work backward in time, building up an an experimental policy that might be better. Suppose estimate of the future value of being in different in- the experimental policy is unlikely to produce persist- formation states p(t) while asking at each time what ent effects (over > 1 yr) in any unit where it is applied, the optimum n is in relation to the current state and and that analysis of its possible performance relative to the baseline has resulted in the following decision that might result from the decision at time t. A sample table of possible outcomes: computation of the optimum relationship between n Hypotheses Policies Conservative Optimistic expected future values of the possible states p(t + 1) and p(t) is shown in Fig. 1, for the case N = 10 and a normal likelihood function with variance 0.25 around mean performance for each unit. 1.0 Fig. 1 shows that management prescriptions can be Experimental 0.5 1.2 very different from scientific ones. When no future Baseline 1.0 learning is expected, the optimum is either to treat all Here the conservative hypothesis is that the baseline units with the experimental policy (for p < 2/7), or to policy is already optimum, so that experimental disturbance of any unit would on average cause a considerable loss (to 0.5). The optimistic hypothesis is that treat none of them. When learning is considered and the units are assumed to have independent random variation, the optimum is a graded policy that involves This content downloaded from 129.219.247.33 on Tue, 08 Jan 2019 16:56:53 UTC All use subject to https://about.jstor.org/terms 2064 SPECIAL FEATURE Ecology, Vol. 71, No. 6 10 -o\---INDEPENDENT w z 8IN--- CORRELATED o -a---X| NOLEARNING w 6L 0 M 4 -^^ z 0- 0 01 0.0 0.2 0.4 0.6 0.8 1.0 PROBABILITY OF FAILURE FIG. 1. Effect of the probability of failure of experimental policy on the optimum number of management units required for treatment with an experimental policy that would be best if an alternative, optimistic hypothesis is correct. Three scenarios are presented: Independent (units respond independently to environmental fluctuations); Correlated (units respond in concert with environmental fluctuations); and, No Learning (manager chooses experimentation only if current conservative policy has a low probability of being correct). treating all the units when p(t) is less than -0.45, and is a long history of sad experience with the false premise to treat progressively fewer units (take fewer risks, do that it is possible to "learn by doing" through sequen- a smaller experiment) as p(t) increases. When the units tial application of different policies to whole systems, are assumed to have correlated responses to environ- especially in fisheries (Walters and Collie 1988). In mental fluctuations, the optimum is to treat fewer units these cases, there is little prospect of resolving the un- and to give up experimentation completely at a lower certainties through continued monitoring and modest value ofp(t). The optimum scientific design in this case policy change, and policy changes drastic enough to would be to treat n = 5 units every year (balanced provide unequivocal responses would be socially or design), to achieve a minimum variance of the treat- economically unacceptable. ment-control difference. Nowhere in these policies can Like ecological processes, policies act at a variety of we see patterns that can be interpreted in terms of spatial scales. For policies that can be implemented at standard scientific ideas of statistical significance. In this example, we assumed that all costs and ben- relatively small scales on a number of experimental units, the main design problem is to control for larger efits of experimentation, including any monitoring costs, scale biophysical processes and management actions are contained in the simple decision table. A fruitful that may link the units. An extreme example is in the area for future statistical research would be to look Columbia River Basin, where the Northwest Power more carefully at the cost structure, and particularly at Planning Council is coordinating development of a how increased monitoring investment would affect the massive ($ 100 million/yr) adaptive management pro- likelihood function L(t), and hence, learning rates. We gram for mitigating effects of water development on suspect that the result of this analysis would most often anadromous salmonid populations (Lee 1989). These be to recommend larger (larger N, n) experiments and populations are linked through factors ranging from more crude monitoring, rather than the precise and large-scale marine and freshwater climate variation to small experiments usually favored by ecological re- shared difficulties in migrating past lower river dams searchers. to mixed-stock fisheries that intercept whole collecSCALE AND REPLICATION Adaptive policy design must make effective use of opportunities for spatial replication and control. There tions of migrating populations. Some policy changes, such as reductions in marine harvest rates and improved fish passage facilities at lower river dams, cannot be replicated in space and will affect any localized This content downloaded from 129.219.247.33 on Tue, 08 Jan 2019 16:56:53 UTC All use subject to https://about.jstor.org/terms December 1990 LARGE-SCALE PERTURBATIONS 2065 experiments within the Basin. Other changes, such as populations through a variety of treatments ranging large hatchery developments and improved passage at from habitat improvement to hatchery rearing. The upriver dams, will affect large segments of the Basin program was divided into two phases, with the first but can be replicated at least a few times. Still others, phase (10 yr) to involve mainly pilot studies (experi- such as habitat improvements in small streams, can ments) with alternative treatments, and the second to be replicated massively throughout the Basin (and in- involve implementation of the best treatments found cidentally, provide a wealth of basic scientific infor- in the first phase. Due to engineering and economic mation on the mechanisms of population regulation in cost-benefit considerations, the bulk of the first phase stream fishes). The challenge is to develop a nested investment went to a set of large hatcheries with an experimental design that will permit clear separation admirable commitment to monitor post-release sur- of the effects of as many of these changes as possible, vival rates through coded wire tagging of juveniles. At so that a sensible balance of management tools and first, the hatcheries performed well, but later there was policies can be developed. THE PROBLEM OF TRANSIENT RESPONSES Most actions do not simply change a managed sys- a massive drop in marine survival rates from several facilities, especially for chinook salmon, and low sur- vival rates have persisted to the present. SEP biologists have contended that the hatchery stocks are not failing, tem from one state to another; rather, they induce tran- but rather that all of the juvenile salmon have been sient responses that may be quite complex (delays, sharp adversely affected by warm water conditions that have increases followed by slow decline, cycles, etc.). More prevailed in the North Pacific since the late 1970s. important in terms of experimental design, some ac- When confronted with evidence that wild ("control") tions may change the sensitivity of managed systems salmon stocks have not shown consistent survival re- to natural environmental factors that themselves have ductions in the same marine environment over the complex temporal patterns. In statistical terms, this same period (Walters and Riddell 1986), SEP propo- change in sensitivity will result in "time-treatment in- nents have countered with an argument involving time- teractions," where response to treatment depends on treatment interaction: treated (hatchery) populations the specific time (environmental conditions) when the may be more sensitive to marine conditions than con- treatment is applied (Walters et al. 1988, 1989). trol (wild) populations, due to some mechanism such The major implication of time-treatment interaction as higher prevalence among hatchery fish of diseases is that treatment-reference comparisons cannot be whose expression is temperature dependent. This apol- trusted to provide a reliable estimate of the treatment ogy is a "there exists" hypothesis that cannot be re- transient effect, no matter how many treatment and jected by detailed research showing that particular reference replicates are used, unless treatment is ini- mechanisms (e.g., diseases) are not responsible for the tiated over a range of starting times in what we (Walters decline. The only way to reject it is to start more hatch- et al. 1988) have called a "staircase" experimental de- eries later in time, then demonstrate that they undergo sign. Such designs require far more experimental units than would ordinarily be considered practical in field would obviously require far longer (_25 yr) than the studies. decade planned for experimentation in the SEP. As an the same transient decline. This experimental design It might be acceptable to ignore time-treatment in- aside, a few SEP hatcheries have come on-line during teractions in basic research experiments. By initiating the late 1970s and early 1980s, and appear to be fol- all treatment and reference comparisons at a single lowing the same transient of high to low survival as moment, one accepts a risk that results may be due in the earlier facilities. part to the specific time one chose to start the exper- iment, so that others may have difficulty in reproducing them. In management settings, time-treatment inter- WHERE TO INVEST EFFORT actions have a much more damaging effect, since they Where replication is possible in principle, we see two can be used by proponents of any management regime main challenges in the design of large-scale manage- to explain away any apparent failures of that regime ment experiments. First, there is a critical need for and hence delay (often at great cost) the implementa- research on imaginative ways to set priorities for in- tion of other options. vesting in research, monitoring, and management. We A costly example of the danger of ignoring time- must find substitutes for many of the cumbersome, treatment interactions in management experiments has time- and worker-intensive sampling methods used in been in hatchery rearing of Pacific salmon in the Canadian Salmonid Enhancement Program (SEP). When most ecological field research, even if the substitutes this $300 million program was initiated in 1974, it was billed as an experimental program to increase salmonid niques such as satellite image analysis and digital par- involve substantial loss in sampling precision. Tech- ticle counting should be seen as more than labor-saving This content downloaded from 129.219.247.33 on Tue, 08 Jan 2019 16:56:53 UTC All use subject to https://about.jstor.org/terms 2066 SPECIAL FEATURE conveniences; they are central to the future of field Ecology, Vol. 71, No. 6 subsequent methodological development (Holling 1978) and process studies (Holling and Buckingham 1976). experimentation. Second, we need institutional arrangements that will However, there are still disagreements concerning the permit and foster experimental studies that span time underlying causes of the time behavior (see Clark et scales longer than the working lives of the scientists al. 1979 vs. Royama 1984). In this example, as well who initiate them. In government management agen- as many others (e.g., fire management, Christensen et cies it is possible to bureaucratize experimental pro- al. 1989), sudden surprises and unexpected behaviors grams to the point where they may even outlive their challenged traditional myths of causation and man- usefulness. The more difficult challenge is to develop agement. During the 1960s, for example, outbreaks incentive systems that will encourage persistent in- began in Newfoundland where the insect was histori- volvement of researchers in the design and conduct of cally rare. More recently, they have become more fre- such programs. quent in stands of young trees. Was knowledge deficient Where replication is impossible and the severity of or had the system fundamentally changed because of disturbance experiments is limited by risks of social expansion of the geographical scale of relevant inter- and economic harm, resource managers will perhaps actions as a consequence of human activities? always operate in a twilight of uncertainty about the The expansion of budworm outbreaks to young stands relative importance of their actions as opposed to the has been particularly surprising, and there are at least effects of uncontrolled environmental and ecological two possible explanations for it. One is the increased factors. By careful process research and modelling we incidence of warm, dry springs, perhaps as a conse- may hope to narrow the range of credible hypotheses quence of global warming. The second concerns changes for the patterns that are seen (Holling 1988), and we in the densities of insectivorous birds. There are about can expect better statistical tools for deciding whether 35 species, most of which migrate to overwintering significant changes are occurring over time (Carpenter areas in the subtropics and tropics where they may be et al. 1989, Goldman et al. 1989). But we must not impacted by deforestation. Is it possible that the mi- pretend that process research and diligent data analysis gration of the birds connects impacts of deforestation alone will provide answers that resource managers can in the tropics with unexpected outbreaks of insects in trust. coniferous forests of North America? Both explana- A critical antecedent to the use of adaptive manage- tions are defensible, but are they credible enough to ment experiments and to decisions of where to invest justify major changes in research and management in- effort is a small number of credible hypotheses to ex- vestments? plain the patterns perceived. It is a trivial task to define The insectivorous bird hypothesis has been explored testable hypotheses, but it is not easy to generate hy- in detail (Holling 1988). The question had to be posed potheses that are relevant to changes in the external in a qualitative way; i.e., how much would bird pop- context and internal structure of managed ecosystems. ulations have to be reduced in order to change the These changes are the real source of the surprises and qualitative behavior of the system? The answer, based crises that pace learning. By focusing on the causes of on existing budworm population models, was by more such abrupt and unexpected changes in behavior, how- than two-thirds, and such a reduction would likely in- ever, it becomes possible to use modelling to evaluate volve even more dramatic reductions in some bird existing process knowledge so as to screen the credible species. It is highly unlikely that such a dramatic change hypotheses, and identify where to concentrate scarce would go undetected by the various professional and amateur bird censuses. Hence, it is not credible that resources. The regional impacts of spruce budworm (Choris- declines in populations of migratory birds are the prox- toneurafumiferana) on forests of eastern North Amer- imate and exclusive cause of outbreaks of budworm in ica provide a typical example. Prior to management, stands of young trees. the insect periodically caused extensive mortality to The identification and evaluation of the set of pos- balsam fir over large areas in eastern Canada and the sibilities discussed in this example depends upon qual- New England states. In New Brunswick, the outbreaks itative analyses of a suite of models and key processes. occurred roughly every 40 yr, causing up to 80% mor- Not one of those models covers all the range of scales tality in mature balsam fir stands. Stands of young trees rarely experienced damage. Since the early 1950s, im- that are relevant. The methods are not available to do so. There are detailed, spatially explicit models that pacts have been intensively managed by insecticide simulate regional budworm-forest dynamics and deal spraying. Coincident with this program was a research explicitly with moth dispersal over a 70 000 km2 area project that still stands as a classic example of inter- (Clark et al. 1979). Other models represent local sites disciplinary analysis of a large-scale ecological system (stands) and were designed to simplify optimization (Morris 1963), and provided the foundation for much studies (Holling et al. 1986) or qualitative stability This content downloaded from 129.219.247.33 on Tue, 08 Jan 2019 16:56:53 UTC All use subject to https://about.jstor.org/terms December 1990 LARGE-SCALE PERTURBATIONS 2067 analysis (Ludwig et al. 1978). Each of these scale-con- ence and politics of the Greenhouse Gas Effect. Both strained models, however, together allowed the eval- the science of parts (e.g., the role of phosphates in lakes) uation of cross-scale interactions by focusing addition- and the science of integration play roles (e.g., global al knowledge of local dispersal of larvae, fragmentation circulation models). In both cases, decisions are not of the landscape and hemispheric movements of birds made because of a well-proofed argument in the tra- and global changes in climate. The range of defensible hypotheses in the budworm dition of experimental science, but because of the accumulation of credible evidence supporting a simple example is typical of many current issues. What is and widely perceived explanation in a political envi- needed are techniques to concentrate on credible pos- ronment that demands action. Hence, resource policy sibilities and structure their evaluation. A blend of decisions can be facilitated by explicit ways to identify scale-constrained models, scale-unconstrained process alternatives, their likelihood and their outcomes in an knowledge, good old-fashioned natural history and ac- environment that engages science, government, and tive adaptive management provides a fruitful direction the public. for both the science and the management of regional renewable resource systems. CONCLUSIONS Two kinds of science influence renewable resource When policies are defined, management begins and the same process of design and analysis occurs, but now in an environment where action has to be taken, however uncertain the outcome. That is where active adaptive management can play a central role, because policy and management. One is a science of parts, e.g., its premise is that knowledge of the system we deal analysis of specific biophysical processes that affect sur- with is always incomplete. Not only is the science in- vival, growth, and dispersal of target variables. It complete, the system itself is a moving target, evolving emerges from traditions of experimental science where because of the impacts of management and the pro- a narrow enough focus is chosen in order to develop gressive expansion of the scale of human influences on data and critical tests that will reject invalid hypoth- the planet. Hence, the actions needed by management eses. The goal is to narrow uncertainty to the point must be ones that achieve ever-changing understanding where acceptance of an argument among scientific peers as well as the social goals desired. That is the heart of is essentially unanimous. It is appropriately conser- active experimentation at the scales appropriate to the vative, unambiguous, and incomplete. The other is a question. Otherwise the pathologies of management science of the integration of parts. It uses the results are inevitable-increasingly fragile systems, myopic of the first, but identifies gaps, invents alternatives, and management, and social dependencies leading to crises evaluates the integrated consequence against planned (Holling 1986). and unplanned interventions in the whole system that occurs in nature. Typically, alternative hypotheses are developed concerning the integrated properties of the whole to reveal the simple causation that often underlies the time and space dynamics of complex systems. Often there is more concern that a useful hypothesis will be rejected than a false one accepted; "don't throw out the baby with the bath water." Since uncertainty is high, the analysis of uncertainty becomes a topic in itself. Renewable resource policy and management has its own profile for achieving degrees of integration and accepting degrees of uncertainty. The degree of match with the science is a measure of the usefulness of the science for decision and action. Policy is politics. Design and acceptance of policy occurs when simple explanations of the causes of and solutions to problems of economic or social consequence achieve sufficient credibility in scientific, government, and public communities. Scientific uncertainty can be high so long as acceptability is high. Remember the phosphate debate of the '60s, when Lake Erie was "dying," that led to the Great Lakes Water Quality Agreement. Or think of the present sci- LITERATURE CITED Anderson, D. R. 1975. Optimal exploitation strategies for an animal population in a Markovian environment: a theory and an example. Ecology 56:1281-1297. Carpenter, S. R., T. M. Frost, D. Heisey, and T. K. Kratz. 1989. Randomized intervention analysis and the interpretation of whole-ecosystem experiments. Ecology 70: 1142-1152. Christensen, N. L., J. K. Agee, P. F. Brussard, J. Hughes, D. H. Knight, G. W. Minshall, J. M. Peek, S. J. Pyne, F. J. Swanson, J. W. Thomas, S. Wells, S. E. Williams, and H. A. Wright. 1989. Interpreting the Yellowstone fires of 1988. BioScience 10:678-685. Clark, W. C., D. D. Jones, and C. S. Holling. 1979. Lessons for ecological policy design: a case study of ecosystem management. Ecological Modelling 7:1-53. Collie, J. S., and C. J. Walters. 1987. Alternative recruitment models of Adams River sockeye salmon (Oncorhynchus nerka). Canadian Journal of Fisheries and Aquatic Sciences 44:1551-1561. DeGroot, M. H. 1970. Optimal statistical decisions. McGraw-Hill, New York, New York, USA. Frederick, P. C., and M. W. Collopy. 1988. Reproductive ecology of wading birds in relation to water conditions in the Florida Everglades. Florida Cooperative Fish and Wildlife Research Unit, School of Forestry Research and Conservation, University of Florida, Technical Report Number 30. This content downloaded from 129.219.247.33 on Tue, 08 Jan 2019 16:56:53 UTC All use subject to https://about.jstor.org/terms 2068 SPECIAL FEATURE Ecology, Vol. 71, No. 6 Goldman, C. R., A. Jassby, andT. Powell. 1989. Interannual Mangel, M., and C. W. Clark. 1988. Dynamic modelling in fluctuations in primary production: meteorological forcing behavioral ecology. Princeton University Press, Princeton, New Jersey, USA. Morris, R. F., editor. 1963. The dynamics of epidemic spruce at two subalpine lakes. Limnology and Oceanography 34: 310-323. Holling, C. S., editor. 1978. Adaptive environmental as- sessment and management. John Wiley and Sons, London, England. 1986. The resilience of terrestrial ecosystems; local surprise and global change. Pages 292-317 in W. C. Clark and R. E. Munn, editors. Sustainable development of the biosphere. Cambridge University Press, Cambridge, England. . 1988. Temperate forest insect outbreaks, tropical deforestation, and migratory birds. Memoirs of the Entomological Society of Canada 146:21-32. Holling, C. S., and S. Buckingham. 1976. A behavioral model of predator-prey functional responses. Behavioral Science 3:183-195. Holling, C. S., G. B. Dantzig, and C. Winkler. 1986. Determining optimal policies for ecosystems. Pages 453-473 in M. Kallio, A. E. Andersson, R. Seppala, and A. Morgan, editors. Systems analysis in forestry and forest industries. Volume 21. TIMS studies in the management science. NorthHolland, Amsterdam, The Netherlands. Keeney, R. L., and H. Raiffa. 1976. Decisions with multiple objectives. John Wiley, New York, New York, USA. Kushlan, J. A. 1987. External threats and internal management: the hydrologic regulation of the Everglades, Florida, USA. Environmental Management 11: 109-119. Kushlan, J. A., and P. C. Frohring. 1986. The history of the southern Florida Wood Stork population. Wilson Bulletin 98:368-386. Lee, K. N. 1989. The Columbia River Basin: experimenting with sustainability. Environment 31(6):7-11, 30-33. Lindley, D. V. 1985. Making decisions. John Wiley, New York, New York, USA. Ludwig, D., D. D. Jones, and C. S. Holling. 1978. Qualitative analysis of insect outbreak systems: the spruce budworm and forest. Journal of Animal Ecology 44:315-332. Mangel, M. 1985. Decision and control in uncertain resource systems. Academic Press, New York, New York, USA. budworm populations. Memoirs of the Entomological Society of Canada 31. Ogden, J. C. 1978. Recent population trends of colonial wading birds on the Atlantic and Gulf coastal plains. In A. Sprunt IV, J. C. Ogden, and S. Winkler, editors. Wading birds. National Audubon Society Research Report 7:135153. Raiffa, H. 1968. Decision analysis: introductory lectures on choices under uncertainty. Addison-Wesley, Reading, Pennsylvania, USA. Royama, T. 1984. Population dynamics of the spruce budworm Choristoneura fumiferana. Ecological Monographs 54:429-462. Walters, C. J. 1981. Optimum escapements in the face of alternative recruitment hypotheses. Canadian Journal of Fisheries and Aquatic Sciences 38:704-710. 1986. Adaptive management of renewable resources. McGraw-Hill, New York, New York, USA. 1990. Value of short term recruitment forecasts for harvest management. Canadian Journal of Fisheries and Aquatic Sciences 47, in press. Walters, C. J., and J. S. Collie. 1988. Is research on environmental effects on recruitment worthwhile? Canadian Journal of Fisheries and Aquatic Sciences 45:1848-1854. Walters, C. J., J. S. Collie, and T. Webb. 1988. Experimental designs for estimating transient responses to management disturbances. Canadian Journal of Fisheries and Aquatic Sciences 45:530-538. Walters, C. J., J. S. Collie, and T. Webb. 1989. Experimental designs for estimating transient responses to habitat alterations: is it practical to control for environmental changes? Canadian Special Publication of Fisheries and Aquatic Sciences, Habitat Symposium, in press. Walters, C. J., and B. Riddell. 1986. Multiple objectives in salmon management: the chinook sport fishery in the Strait of Georgia, British Columbia. Northwest Environmental Journal 2:1-15. This content downloaded from 129.219.247.33 on Tue, 08 Jan 2019 16:56:53 UTC All use subject to https://about.jstor.org/terms