Frequency selection in paleoclimate time series: A model-based approach incorporating possible time uncertainty
Frequency selection in paleoclimate time series: A model-based approach incorporating possible time uncertainty
复制标题
古气候时间序列中的频率选择：基于模型的方法，考虑了可能的时间不确定性
DOI：
10.1002/env.2492
复制
发表时间：
2018
期刊：
影响因子：
1.7
通讯作者：
Franke P
中科院分区：
文献类型：
--
作者：
Franke P
古气候时间序列分析的一个关键方面是频率行为的识别。通常，这是通过计算功率谱并将该谱与简化模型的谱进行比较来实现的。传统的假设检验方法，然后可以用来找到统计上显着的峰值，对应于不同的频率。当数据是多变量的或具有时间不确定性时，会出现复杂性。特别是，联合周围的不确定性观测和他们的时间的存在，使传统的假设testingpracticed.In本文中，我们重新表达的频率识别问题在时域中的变量选择模型，每个变量对应于一个不同的频率。我们把这个问题的贝叶斯框架，使我们能够把收缩先验分布的权重，每个频率，以及包括信息丰富的先验信息，通过它，我们可以考虑到时间的不确定性，我们验证我们的方法与模拟数据，并说明它与分析中晚全新世水位记录从北方爱尔兰死岛和Slieveanorra。这两个案例研究也显示了研究人员可能面临的挑战的程度。因此，我们提出了一种情况下，显示了一个很好的模型拟合一个明确的频率模式和其他情况下，频率行为的识别是不可能的。我们将我们的结果与现存的方法，称为REDFIT。
A key aspect of paleoclimate time series analysis is the identification of frequency behavior. Commonly, this is achieved by calculating a power spectrum and comparing this spectrum with that of a simplified model. Traditional hypothesis testing method can then be used to find statistically significant peaks that correspond to different frequencies. Complications occur when the data are multivariate or suffer from time uncertainty. In particular, the presence of joint uncertainties surrounding observations and their timing makes traditional hypothesis testing impractical.In this paper, we reexpress the frequency identification problem in the time domain as a variable selection model where each variable corresponds to a different frequency. We place this problem in a Bayesian framework that allows us to place shrinkage prior distributions on the weighting of each frequency, as well as include informative prior information through which we can take account of time uncertainty.We validate our approach with simulated data and illustrate it with analysis of mid‐ to late Holocene water table records from two sites in Northern Ireland—Dead Island and Slieveanorra. Both case studies also show the extent of the challenges that researchers may face. We therefore present one case that shows a good model fit with a clear frequency pattern and the other case where the identification of frequency behavior is impossible. We contrast our results with that of the extant methodology, known as REDFIT.