An adaptive composite density estimator for k-tree sampling
An adaptive composite density estimator for k-tree sampling
复制标题
DOI：
10.1007/s10342-011-0502-8
复制
发表时间：
2012-03
影响因子：
2.8
通讯作者：
S. Magnussen;L. Fehrman;W. Platt
中科院分区：
文献类型：
--
作者：
S. Magnussen;L. Fehrman;W. Platt
叉树距离抽样的密度估计量对到第k棵树的距离的额外泊松方差的量敏感。为了减少这种敏感性，我们提出了一个自适应复合估计（COM）。在模拟抽样从16个测试人口，三个组件的复合密度估计（COM）-与权重确定的多项式逻辑函数的四个现成的辅助变量-被确定为上级的平均相对绝对偏差。结果从一个不同的一组9验证人口与广泛不同的干密度和空间格局的树的位置证实，COM的相对均方根误差（RRMSE），平均而言，大大低于与三个组件k-树密度估计。COM的RRMSE性能随着k值的增加而提高。在k = 6和样本量为10、20和30的情况下，COM的平均相对偏差在7个验证群体中介于-5至5%之间，但在开放的低密度大草原样群体中，偏差达到-12%（1979年数据）和7%（1996年数据）。Fork= 6和n = 10，COM的RRMSE是，在9个验证群体中的6个，在3.3个百分点的RRMSE固定面积图抽样。Jackknife估计的COM估计密度的精度是负偏的，导致计算的95%置信区间的覆盖率不足（7%）。
Density estimators fork-tree distance sampling are sensitive to the amount of extra Poisson variance in distances to thekth tree. To lessen this sensitivity, we propose an adaptive composite estimator (COM). In simulated sampling from 16 test populations, a three-component composite density estimator (COM)–with weights determined by a multinomial logistic function of four readily available ancillary variables–was identified as superior in terms of average relative absolute bias. Results from a different set of nine validation populations–with widely different stem densities and spatial patterns of tree locations—confirmed that relative root mean squared errors (RRMSE) of COM were, on average, considerably lower than those obtained with the three-componentk-tree density estimators. The RRMSE performance of COM improved with increasing values ofk. Withk= 6 and sample sizes of 10, 20, and 30, the average relative bias of COM was between −5 and 5% in seven validation populations but in an open low-density savanna-like population bias reached −12% (1979 data) and 7% (1996 data). Fork= 6 andn= 10, the RRMSE of COM was, in six of the nine validation populations, within 3.3 percentage points of the RRMSE for sampling with fixed-area plots. Jackknife estimates of the precision of COM estimates of density were negatively biased, leading to under-coverage (7%) of computed 95% confidence intervals.