Numerous statistical models are employed when processing forest inventory data. These models primarily provide predicted values for attributes that are difficult and/or time-consuming to measure. In some applications, models are applied across a large geographic area, which assumes the relationship between the response variable and predictors is constant within the area. The extent to which this assumption holds for a tree height prediction model was evaluated at regional, ecoprovince and

more »

... tion scales in the northeastern US. Two nonlinear regression models were tested, a spatially ambiguous model that utilized tree and stand-level predictors, and a spatially explicit model that incorporated latitude, longitude and elevation as predictors. When the regional-scale models were evaluated at the state level, both showed considerable bias for some states, which suggests that the statistical significance of spatial predictor variables does not translate into effective accounting for spatial variability. Similar results were obtained when fitting the model to an ecoprovince and evaluating bias within ecosections. Finally, fitting the model to ecosections within the ecoprovince provided a moderate level of local robustness as assessed by Moran's I statistic; however, there are cases where local biases may still exist. This outcome suggests that models should be developed and applied at small spatial scales to reduce local biases when model predictions are aggregated to larger geographic domains. Alternatively, more advanced modelling techniques may be more effective at addressing local variability using a single model having large-area application. However, the practicality of implementing these more complex techniques in the context of continuous large-area forest inventories is not well understood and should be fully explored prior to operational employment.