II. APPLICABILITY OF PREDICTIVE (FORECASTING) MODELS 



The utility of a model to a manager depends on the amount of confidence 

 that can be assigned to its predictions or forecasts. This "level of confidence" 

 is influenced by many factors which interact synergist ically to enhance or 

 diminish the predictive capability (utility) of a model. Basic among these 

 factors are the nature of the management/assessment objective, the data base, 

 and the degree of ecological complexity which must be modeled to minimally 

 address the resource management and impact assessment concerns. In general, 

 models addressing problems that involve few, noninteracting species and processes 

 work best. 



The properties and dynamics of seawater are far better understood than 

 those of marine populations, communities, or ecosystems. Thus, the more a 

 problem is dominated by deterministic physical processes, the better it can be 

 modeled. Examples are nearshore spill and plume trajectories and settling of 

 dredging spoils. Conversely, the more a problem is dominated by complex and 

 interactive biological processes, the less easily it can be modeled. Uncertain- 

 ties about the biological structure, function, and interrelationships of a 

 population, community, or ecosystem usually overwhelm all other factors that 

 might limit forecasting ability. Also, as the time and space scales and 

 resolution of the management problem expand, the predictive capability diminishes, 

 Again, this most strongly affects the biological models. For example, the 

 trajectories of nearshore spills and plumes can be predicted with reasonable 

 accuracy for 1 to 2 weeks into the future. However, as the temporal and spatial 

 boundaries of the problem expand, weather-related chance events and uncertainties 

 about the physiology of species, the life histories of populations, the structure 

 and function of communities , interactions among communities , and rates of biotic 

 energy and material flows become overwhelming. As a result, short-term (day, 

 week) acute impacts such as mortalities in habitats contacted by high concentra- 

 tions of toxic materials can be modeled with much greater predictive success/ 

 confidence than can long-term (month, year) chronic impacts such as physiological 

 and genetic damage caused by constant low or high toxicant concentrations. 



Process dynamics (e.g., primary production, grazing, nitrification) and 

 single species population dynamics can be modeled with greater confidence than 

 community dynamics. Empirical processes and population models (e.g., fisheries 

 stock models) that essentially interpolate data have consistently yielded 

 reliable predictions as long as interpolation limits are not exceeded; they 

 quickly break down outside of those limits. Mechanistic processes models (e.g., 

 uptake and grazing kinetics) not only can interpolate data, but they can also, 

 at least in theory, extrapolate beyond the ranges of experience. A critical 

 factor to the predictive success of both approaches is the data record upon 

 which the conceptual and mathematical models are based. The least squares, 

 best fit methodology of the empirical models requires an extensive data base of 

 field observations; a length of record of 20 yr is often the preferred minimum. 

 In addition, as the empirical relationships deviate further from linearity 

 there is a concomitant need for greater temporal and spatial resolution in the 

 data base. Because of their cause and effect approach, mechanistic models are 

 not as dependent on lengthy records of field observations and field data 

 resolution as are the empirical models. Mechanistic models can derive/ formulate 

 causal relationships from laboratory experiments. However, a degree of 

 uncertainty is always introduced when laboratory-derived relationships are 

 applied to the natural ecosystem. This uncertainty becomes magnified when the 



258 



