Strategies for the use of mixed-effects models in continuous forest inventories

Abstract

Forest inventory data often consists of measurements taken on field plots as well as values predicted from statistical models, e.g., tree biomass. Many of these models only include fixed-effects parameters either because at the time the models were established, mixed-effects model theory had not yet been thoroughly developed or the use of mixed models was deemed unnecessary or too complex. Over the last two decades, considerable research has been conducted on the use of mixed models in forestry, such that mixed models and their applications are generally well understood. However, most of these assessments have focused on static validation data, and mixed model applications in the context of continuous forest inventories have not been evaluated. In comparison to fixed-effects models, the results of this study showed that mixed models can provide considerable reductions in prediction bias and variance for the population and also for subpopulations therein. However, the random effects resulting from the initial model fit deteriorated rapidly over time, such that some field data is needed to effectively recalibrate the random effects for each inventory cycle. Thus, implementation of mixed models requires ongoing maintenance to reap the benefits of improved predictive behavior. Forest inventory managers must determine if this gain in predictive power outweighs the additional effort needed to employ mixed models in a temporal framework.