Evaluating alternative implementations of the Lake States FVS diameter increment model

We evaluated the STEMS family of diameter increment models that have been incorporated in the Lake States variant of the Forest Vegetation Simulator. We paired validation using regression-based equivalence tests with evaluation of trends in errors across species and predictor variables, using independent data from the Michigan Forest Inventory and Analysis program. Our evaluation shows that 10-year increment bias is substantial, almost 17% on average, and our tests failed to validate the model for every one of the 30 most common tree species in the region. A comparative analysis among all alternative implementations demonstrated that error arose from structural weaknesses in the underlying model. Furthermore, the way the model is currently implemented in the Forest Vegetation Simulator partly masks poor performance at the tree level, but likely amplifies error at the stand level, a particularly troubling result in many conceivable applications. Our results also affirm that a simple adjustment factor as a function of dbh provides an inadequate correction of prediction bias. We argue that the diameter increment model needs to be re-engineered.