Bryssinckx Ward, Ducheyne Els, Leirs Herwig, Hendrickx Guy
Geospat Health. 2014 May;8(2):573-82. doi: 10.4081/gh.2014.47.
To assess the risk of (zoonotic) disease transmission in developing countries, decision makers generally rely on distribution estimates of animals from survey records or projections of historical enumeration results. Given the high cost of large-scale surveys, the sample size is often restricted and the accuracy of estimates is therefore low, especially when spatial high-resolution is applied. This study explores possibilities of improving the accuracy of livestock distribution maps without additional samples using spatial modelling based on regression tree forest models, developed using subsets of the Uganda 2008 Livestock Census data, and several covariates. The accuracy of these spatial models as well as the accuracy of an ensemble of a spatial model and direct estimate was compared to direct estimates and "true" livestock figures based on the entire dataset. The new approach is shown to effectively increase the livestock estimate accuracy (median relative error decrease of 0.166-0.037 for total sample sizes of 80-1,600 animals, respectively). This outcome suggests that the accuracy levels obtained with direct estimates can indeed be achieved with lower sample sizes and the multimodel approach presented here, indicating a more efficient use of financial resources.
为评估发展中国家（人畜共患）疾病传播的风险，决策者通常依赖于根据调查记录得出的动物分布估计数或历史普查结果的预测。鉴于大规模调查成本高昂，样本量往往受限，因此估计的准确性较低，尤其是在应用空间高分辨率数据时。本研究探讨了在不增加样本的情况下提高牲畜分布图准确性的可能性，采用基于回归树森林模型的空间建模方法，该模型利用乌干达2008年牲畜普查数据的子集和若干协变量开发而成。将这些空间模型的准确性以及空间模型与直接估计相结合的准确性，与基于整个数据集的直接估计和“真实”牲畜数量进行了比较。新方法被证明能有效提高牲畜估计的准确性（对于样本总量分别为80 - 1600头动物的情况，中位数相对误差从0.166降至0.037）。这一结果表明，使用直接估计获得的准确性水平确实可以通过更小的样本量和本文提出的多模型方法来实现，这意味着能更有效地利用财政资源。
