Department of Ecology, 310 Lewis Hall, Montana State University, Bozeman, Montana 59717, USA.
Ecol Appl. 2013 Apr;23(3):643-53. doi: 10.1890/12-0959.1.
It is increasingly common for studies of animal ecology to use model-based predictions of environmental variables as explanatory or predictor variables, even though model prediction uncertainty is typically unknown. To demonstrate the potential for misleading inferences when model predictions with error are used in place of direct measurements, we compared snow water equivalent (SWE) and snow depth as predicted by the Snow Data Assimilation System (SNODAS) to field measurements of SWE and snow depth. We examined locations on elk (Cervus canadensis) winter ranges in western Wyoming, because modeled data such as SNODAS output are often used for inferences on elk ecology. Overall, SNODAS predictions tended to overestimate field measurements, prediction uncertainty was high, and the difference between SNODAS predictions and field measurements was greater in snow shadows for both snow variables compared to non-snow shadow areas. We used a simple simulation of snow effects on the probability of an elk being killed by a predator to show that, if SNODAS prediction uncertainty was ignored, we might have mistakenly concluded that SWE was not an important factor in where elk were killed in predatory attacks during the winter. In this simulation, we were interested in the effects of snow at finer scales (< 1 km2) than the resolution of SNODAS. If bias were to decrease when SNODAS predictions are averaged over coarser scales, SNODAS would be applicable to population-level ecology studies. In our study, however, averaging predictions over moderate to broad spatial scales (9-2200 km2) did not reduce the differences between SNODAS predictions and field measurements. This study highlights the need to carefully evaluate two issues when using model output as an explanatory variable in subsequent analysis: (1) the model's resolution relative to the scale of the ecological question of interest and (2) the implications of prediction uncertainty on inferences when using model predictions as explanatory or predictor variables.
越来越多的动物生态学研究使用基于模型的环境变量预测作为解释变量或预测变量，尽管模型预测不确定性通常是未知的。为了展示在使用存在误差的模型预测代替直接测量值时进行误导性推断的可能性，我们将 SNODAS 预测的雪水当量 (SWE) 和雪深与 SWE 和雪深的实地测量值进行了比较。我们检查了怀俄明州西部麋鹿 (Cervus canadensis) 冬季牧场的位置，因为像 SNODAS 输出这样的模型数据通常用于麋鹿生态学的推断。总体而言，SNODAS 预测值往往高估了实地测量值，预测不确定性较高，并且与非雪影区域相比，两个雪变量的 SNODAS 预测值与实地测量值之间的差异在雪影区域更大。我们使用一个简单的模拟来展示雪对麋鹿被捕食者杀死的概率的影响，如果忽略 SNODAS 预测不确定性，我们可能错误地得出结论，认为 SWE 不是冬季麋鹿在捕食者攻击中被杀死的重要因素。在这个模拟中，我们对小于 SNODAS 分辨率的更精细尺度 (<1km2) 的雪的影响感兴趣。如果 SNODAS 预测在更粗糙的尺度上平均时偏差减小，那么 SNODAS 将适用于种群水平的生态学研究。然而，在我们的研究中，在中等至广泛的空间尺度 (9-2200km2) 上平均预测值并没有减少 SNODAS 预测值与实地测量值之间的差异。本研究强调了在将模型输出用作后续分析中的解释变量时，需要仔细评估两个问题：(1) 模型相对于感兴趣的生态问题的尺度的分辨率；(2) 在使用模型预测作为解释变量或预测变量时，预测不确定性对推断的影响。
