data at both the physiological and ecosystem level and incorporates a large 

 number of assumptions. The strength, however, is that these models potentially 

 can "predict" the responses of many components of the ecosystem to new or 

 extreme conditions, even when such conditions historically have never existed. 

 These predictions, though they may be important, are based on an explicit series 

 of identifiable statements about the underlying causal mechanisms. When used 

 correctly, with continuous appreciation of the assumptions, these models may be 

 the only quantitative method for evaluating certain environmental scenarios. 



Intermediate Strategies 



In between the extremes of empirical and mechanistic models falls a broad 

 range of approaches, many combining, to various degrees, elements of both. 

 Mechanisms can be specified, but with various theoretical perspectives. For 

 example, many approaches attempt to relate changes in certain state variables 

 to forcing factors in the environment using simple or complex statistical 

 expressions. Thus, linear or nonlinear terms may be used to define the dynamic 

 interactions among the compartments at the ecosystem level. In contrast to the 

 mechanistic strategy, the data required for these formulations of processes are 

 not physiological, but are the synoptic observations of the state variables and 

 forcing functions. Thus, the data requirements are more limited and generally 

 site specific, but again causal mechanisms are less fundamentally represented. 



Deterministic and Stochastic Modeling Approaches 



Output of deterministic models can be seriously misleading if the modeled 

 environment is characterized by variable or uncertain conditions. This is 

 particularly true when both controllable and uncontrollable variables have 

 strong stochastic components. Also, estimates of these future mean properties 

 are often not useful to managers. Quite often managers require probabilities 

 of certain events occurring. For example, the frequency of occurrence and 

 duration of dissolved oxygen (DO) excursions below some acceptable limit could 

 be more important than mean DO concentrations. 



Certain stochastic environmental properties are quantifiable statistically 

 (e.g., current roses, light and temperature variability, and nutrient loadings). 

 We have the means in hand (e.g., Monte Carlo, 1st order variance propagation 

 analyses, random differential equations) to address the effects of this 

 natural variability on model predictions. All models used in predictive or 

 diagnostic modes should take advantage of these or similar methods to account 

 for, or at least address, quantifiable uncertainty. 



Deterministic models of lumped properties (e.g., functional groups of 

 phytoplankton) use parameter values, initial conditions, etc., that presumably 

 represent typical species in that group. However, in nature functional groups 

 are often led by strong dominants whose dynamics may not represent those of the 

 total group. One might expect, though, that the dynamics of individuals within 

 the group (e.g., their kinetic coefficients) may follow some distribution (not 

 necessarily symmetrical) centered on those of the "typical" component. These 

 conditions may be amenable to Monte Carlo analysis. A model solution is 

 obtained many times, each time using kinetic coefficients drawn randomly from 

 some distribution centered on the typical coefficient values. This analysis will 

 produce output that has estimates of both the mean response and spread of 

 responses. While the central tendency may represent the response of the typical 



246 



