Evaluating predictive errors of a complex environmental model using a general linear model and least square means

• Published in: Ecological modelling 2005-08, Vol.186 (3), p.366-374
• Main Authors: Knightes, Christopher D., Cyterski, Michael
• Format: Article
• Language: English
• Subjects: Animal, plant and microbial ecology; Biological and medical sciences; Environmental; Errors; Evaluation; Freshwater; Fundamental and applied biological sciences. Psychology; General aspects. Techniques; General linear model; Methods and techniques (sampling, tagging, trapping, modelling...)
• ISSN: 0304-3800; 1872-7026
• DOI: 10.1016/j.ecolmodel.2005.01.034

A general linear model (GLM) was used to evaluate the deviation of predicted values from expected values for a complex environmental model. For this demonstration, we used the default level interface of the regional mercury cycling model (R-MCM) to simulate epilimnetic total mercury concentrations in Vermont and New Hampshire lakes based on data gathered through the EPAs Regional Environmental Monitoring and Assessment Program (REMAP). The response variable for the GLM was defined as R-MCMs predictive error: the difference between observed mercury concentrations and modeled mercury concentrations in each lake. Least square means of the response variable are used as an estimate of the magnitude and significance of bias, i.e., a statistically discernable trend in predictive errors for a given lake type, e.g., acidic, stratified, or oligotrophic. Using our approach, we determined lake types where significant over-prediction and under-prediction of epilimnetic total mercury concentration was occurring, i.e., regions in parameter space where the model demonstrated significant bias was distinguished from regions where no significant bias existed. This technique is most effective for finding regions of parameter space where bias is significant. Drawing conclusions concerning regions that show no significant bias can be misleading. The significant interaction terms in the GLM demonstrated that addressing this problem using univariate statistical techniques would lead to a loss of important information.