Parametric bootstrap estimators for hybrid inference in forest inventories. (22nd November 2017)

• Record Type: Journal Article
• Authors: Fortin, Mathieu; Manso, Rubén; Schneider, Robert
• Is Part Of: Forestry. Volume 91:Number 3(2018)
• Page Start: 354
• Page End: 365
• Publication Date: 2017-11-22
• Subjects: Forests and forestry -- Periodicals; Forests and forestry -- Great Britain -- Periodicals; 634.9
• DOI: 10.1093/forestry/cpx048
• Languages: English
• ISSNs: 0015-752X

Abstract

Abstract: In forestry, the variable of interest is not always directly available from forest inventories. Consequently, practitioners have to rely on models to obtain predictions of this variable of interest. This context leads to hybrid inference, which is based on both the probability design and the model. Unfortunately, the current analytical hybrid estimators for the variance of the point estimator are mainly based on linear or nonlinear models and their use is limited when the model reaches a high level of complexity. An alternative consists of using a variance estimator based on resampling methods (Rubin, D. B. (1987). Multiple imputation for nonresponse surveys . John Wiley & Sons, Hoboken, New Jersey, USA). However, it turns out that a parametric bootstrap (BS) estimator of the variance can be biased in contexts of hybrid inference. In this study, we designed and tested a corrected BS estimator for the variance of the point estimator, which can easily be implemented as long as all of the stochastic components of the model can be properly simulated. Like previous estimators, this corrected variance estimator also makes it possible to distinguish the contribution of the sampling and the model to the variance of the point estimator. The results of three simulation studies of increasing complexity showed no evidence of bias for this corrected variance estimator, which clearly outperformed the BS variance estimator used in previous studies. Since the implementation of this corrected variance estimator is not much more complicated, we recommend its use in contexts of hybrid inference based on complex models.