Monitoring Is Not Enough: On the Need for a Model-Based Approach to Migratory Bird Management James D. Nichols Abstract—Informed management requires information about system state and about effects of potential management actions on system state. Population monitoring can provide the needed information about system state, as well as information that can be used to investigate effects of management actions. Three methods for investigating effects of management on bird populations are (1) retrospective analysis, (2) formal experimentation and constraineddesign studies, and (3) adaptive management. Retrospective analyses provide weak inferences, regardless of the quality of the monitoring data. The active use of monitoring data in experimental or constrained-design studies or in adaptive management is recommended. Under both approaches, learning occurs via the comparison of estimates from the monitoring program with predictions from competing management models. Monitoring, the activity of making periodic assessments of state variables reflecting system dynamics, has become popular among those studying ecological populations and communities. In the case of bird populations, monitoring programs in North America include efforts to monitor population size (e.g., the North American Breeding Bird Survey [Sauer and Droege 1990], the Audubon Christmas Bird Counts, and numerous other surveys); and the vital rates (survival and reproductive rates) responsible for changes in population size (e.g., the MAPS program [DeSante and others 1995] and BBIRD [Conway and Martin, this proceedings]). When faced with a difficult management decision, decisionmakers frequently exhibit a form of displacement behavior or stalling tactic by asking for additional information, without thinking sufficiently clearly about exactly what kinds of information will be most useful in making the decision. I believe that some monitoring programs have been established as a result of such displacement behavior with the vague hope that future retrospective looks at monitoring results will somehow lead to a reduction in future uncertainty and to an understanding of population dynamics and responses to management. In this paper, I express the view that monitoring can be very useful in helping us to understand and manage bird populations. However, I argue that this utility is dependent upon using monitoring results in conjunction with models In: Bonney, Rick; Pashley, David N.; Cooper, Robert J.; Niles, Larry, eds. 2000. Strategies for bird conservation: The Partners in Flight planning process; Proceedings of the 3rd Partners in Flight Workshop; 1995 October 1-5; Cape May, NJ. Proceedings RMRS-P-16. Ogden, UT: U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station. James D. Nichols, Biological Resources Division, Patuxent Wildlife Research Center, Laurel, MD 20708. USDA Forest Service Proceedings RMRS-P-16. 2000 and hypotheses about system behavior as part of a scientific process such as experimentation or adaptive management (Walters 1986). In the absence of a scientific context, I believe that monitoring, by itself, is not likely to lead to good understanding or management (Nichols 1991a). In this sense, I tend to agree with the following provocative statement by Krebs (1991:3), “Monitoring of populations is politically attractive but ecologically banal unless it is coupled with experimental work to understand the mechanisms behind system changes.” I will try to provide a basis for such views about monitoring and, in the process, identify explicit roles for monitoring data in management and research programs. Management Requisites and the Role of Monitoring ______________ I believe that four fundamental requirements exist for the informed management of any animal population or community (Nichols and others 1995). First, the manager must develop explicit goals for the management process. These goals should be translated into an objective function that specifies the relative values of different specific objectives and the costs of management actions (Nichols and others 1995; Williams 1996; Conroy, this proceedings). Second, the manager must have the ability to implement different management actions relevant to attainment of management objectives. In the example of migratory bird management, relevant actions would most likely involve habitat management, but also could include control of predators or nest parasites. Third, at each decision point, the manager must have access to information about the state of the managed system and about other goal-related variables. For migratory birds, system state would likely include population size(s) and, possibly, related quantities. The fourth requisite for informed management is knowledge about the response of the managed system to alternative management actions. In an uncertain world, this “knowledge” is typically represented by hypotheses or models about the dynamics of the managed system. Monitoring programs are important to the third and fourth management requisites listed above. They are clearly important as the primary source of information about the state of the managed system. In single-species management, we must know something about the population size at time t to make an informed decision about how to manage at that time. Monitoring programs directed at the population variables of interest provide such information for management. The other role of monitoring programs is to provide data that can be used to develop our “knowledge” of system 121 dynamics and responses to management. It is with respect to this acquisition of knowledge that I claim “monitoring is not enough.” Approaches to Learning __________ Retrospective Analyses Monitoring data have been used in various ways to help us understand the managed system. Perhaps the most commonly used approach is the retrospective analysis of timeseries data produced by monitoring programs (Nichols 1991a). Monitoring programs may yield annual estimates of quantities such as population size over some relatively long time period (e.g., 20 years), and it is commonly thought that such trajectories lead almost immediately to an understanding of underlying population dynamics. It is a common practice to use such data with correlation and regression analyses to investigate possible relationships between population size and various environmental and management variables. The problem with this approach is that it is unlikely to yield “reliable knowledge” (Romesburg 1981), because there typically will be multiple a posteriori stories or models that provide reasonable explanations for any observed time series. Indeed, Pirsig (1974:107) suggested that “The number of rational hypotheses that can explain any given phenomenon is infinite.” The potential for being misled by retrospective analysis exists for all kinds of observations (Platt 1964; Romesburg 1981) but is probably especially large for time series of estimates of population size and related variables. One reason for this is that population size is not observed but is estimated, often with large sampling variances and sometimes with bias. Temporal variation in point estimates of population size is thus not equivalent to temporal variation in the underlying population (Link and Nichols 1994). Stories and hypotheses developed to explain the variation in a series of population estimates may not be adequate to explain the real variation in the underlying quantity of interest. Another difficulty in drawing inferences from retrospective analyses of population trajectory data involves the stochastic nature of population processes. Death, for example, is typically viewed as a simple stochastic process. If a population has 100 animals at time t and if each of these animals has a probability of 0.3 of dying during the interval (t, t + 1), then we do not expect exactly 70 animals to be alive at time t + 1. Instead the number of survivors will be a binomial random variable with expected value 70, but with likely realized values of 68, 73, 65, etc. Reproductive processes and movement also are viewed as stochastic in nature, leading to the view of a population trajectory as a single realization of a (likely complicated) stochastic process. We should not expect to be able to observe a single realization of an unknown stochastic process and infer much about the process (Nichols 1991a). This is analogous, in some sense, to being handed a loaded coin, being permitted to flip it once, and then being asked to specify the probability of obtaining heads. Another difficulty associated with inferences from retrospective studies of population monitoring data involves problems with using correlation analysis to draw inferences 122 about the functional relationship between variables represented by time series. A clear example of such problems involves the existence of trends and monotonicity in many environmental covariates that potentially influence bird populations. Metrics of human-related environmental variables such as habitat fragmentation, habitat degradation, and pollutant levels will frequently tend to show an increasing trend over time. Correlation analyses involving two variables, each of which shows a time trend, will tend to indicate association, although this may have nothing to do with any functional relationship between the variables. In fact, the problem of conducting association analyses of two time series extends beyond the case of monotonic trends, and such analyses frequently lead to inappropriate inferences (Yule 1926; Barker and Sauer 1991). The conclusion I draw from this consideration of problems associated with inferences from retrospective studies is that such analyses should not form the basis for our knowledge of population dynamics and responses to management. Instead, I suggest that we restrict our use of retrospective analyses of monitoring data to the generation of hypotheses about population dynamics and responses. We can then use scientific processes such as experimentation and adaptive management to distinguish between competing models. For those who do not find these arguments convincing, I recommend readings on the history of efforts to investigate the effects of hunting regulations on waterfowl populations in North America (Anderson and Burnham 1976; Nichols and others 1984; Nichols 1991b; Smith and Reynolds 1992). These efforts employed retrospective analyses of estimates of population size, survival rates, hunting mortality rates, and reproductive rates produced by some of the best largescale animal monitoring programs in the world, yet have not led to the general acceptance of a single model of waterfowl population dynamics and responses to hunting (Nichols and others 1995). Experiments and Constrained Studies Monitoring data also can be used in conjunction with studies designed specifically to test hypotheses about responses of animal populations to management actions. We begin with competing hypotheses or models about population dynamics and responses to management. We try to discriminate among them by applying different treatments (e.g., different management actions) to birds inhabiting different portions of the landscape. If monitoring programs are in place for the treated areas, then differences in response variables (e.g., population size) between areas treated in different ways may provide the basis for inference about management effects (i.e., may help us decide which model provides the best approximation to reality). Inferences are strongest for true manipulative experiments (Hurlburt 1984; Skalski and Robson 1992). Such experiments are characterized by replication and randomization in the assignment of different treatments to experimental units. Replication refers to the application of treatments to multiple experimental units as a means of estimating the experimental error or error variance (the variance among experimental units to be expected in the absence of treatment differences; i.e., the variance associated USDA Forest Service Proceedings RMRS-P-16. 2000 with all factors except the different treatments). Randomization refers to random assignment of treatments to experimental units. Randomization protects against systematic differences among experimental units receiving different treatments and represents an effort to insure that any systematic post-treatment differences among experimental units treated differently can be attributed to the treatments themselves. In ecological studies in general, and studies of managed populations and communities in particular, manipulative experiments are frequently difficult to perform. In such instances, we may be forced to use study designs that lack replication, randomization, or both of these features. Such constrained designs are considered in some detail by Green (1979) and Skalski and Robson (1992). Inferences resulting from such studies will typically be stronger than those resulting from retrospective analyses, yet not as strong as those based on manipulative experimentation (see examples in Nichols and Johnson 1989). The important point is that studies, whether they are true manipulative experiments or constrained studies, can be designed to take advantage of monitoring programs. We begin with different a priori hypotheses about population dynamics and responses to management. The monitoring programs provide the estimates of potential response variables that are used to distinguish among competing hypotheses, and test for effects of the studied treatments (typically management actions). In such studies the short-term objectives of management are temporarily set aside in an attempt to learn something about the response of the studied system to management practices (see Conroy, this proceedings and Cooper and others, this proceedings for examples of such studies). However, any knowledge of system response will be useful in making subsequent management decisions and, through the reduction of uncertainty, will be beneficial to long-term management objectives. Adaptive Management Adaptive management (Walters 1986) presents another approach to using monitoring data to help us distinguish among competing hypotheses. Under this approach, however, we do not temporarily set aside short-term management objectives to learn something about the managed system. Instead, adaptive management represents an approach to dealing with the so-called “dual control problem” of simultaneously meeting short-term management objectives and learning about the managed system to reduce uncertainty so that future objectives can be better met. Adaptive management requires five components that are closely related to the four requisites for informed management listed earlier in this paper: 1. The objective function is simply a formal specification of management objectives. 2. Management options are actions that are available to the manager and that are potentially relevant to attainment of management objectives. 3. The monitoring program provides information on system state variables and other goal-related variables. USDA Forest Service Proceedings RMRS-P-16. 2000 4. The model set is a set of competing models about system dynamics and response to management actions. 5. A measure of uncertainty is required for each model in the model set. These measures are probabilities that sum to one and reflect our relative faith in the different models. At a decision point, t, the monitoring program provides information about the state of the managed system. Given this state, given the objective function, and given the members of the model set with their respective measures of uncertainty, we derive the management action that is optimal with respect to the weighted (by the measures of uncertainty) predictions of the different models. This action is implemented and drives the system to a different state at the next decision point, t + 1. This new state is estimated via the monitoring program, and this estimate is used to update or revise our measures of uncertainty. If the new state of the system corresponds closely to the prediction of a particular model, then the probability associated with that model is increased. If the new state of the system deviates substantially (relative to other models) from the prediction of a model, then the probability associated with that model is decreased. At time t + 1, the optimal decision is again derived given the new state of the system and the new model weights. The process of adaptive management proceeds in this iterative fashion. Our measures of uncertainty change through time and reflect learning about system responses. This learning leads to a reduction in uncertainty that improves our ability to meet management objectives (Conroy, this proceedings; Nichols and others 1995; Walters 1986; Williams 1996). Conclusions ____________________ Population monitoring has two explicit roles in the process of managing an animal population. First, it provides information about the state of the managed population at each time period for which a management decision is needed. Second, monitoring can contribute to our knowledge of effects of management actions on populations. Three different approaches to the use of monitoring data to aid understanding of animal population dynamics and responses to management were discussed. It was concluded that the most common approach, using monitoring data in retrospective analyses, is not likely to yield “knowledge” in which we can be confident and on which we will be willing to base management decisions. Instead, retrospective analyses of monitoring data can be used to develop hypotheses and models of animal populations and management responses. Monitoring data can then be used in conjunction with either experimental or constrained studies, or adaptive management, to estimate population responses to management. These estimated responses can be compared with predictions of competing models, and results of these predictions can then be used to distinguish among the different models. The key to the use of monitoring data with experimental or constrained studies, or adaptive management, is the existence of a priori hypotheses about system dynamics and response. When the roles of monitoring data are made explicit in this manner, it becomes easier to design monitoring programs to meet management needs. 123