Often power transformations can stabilize variability in variance across a geographic landscape. For example,
这个变换过程就是Box-Cox变换,即Box和Cox在1964年提出的一种广义幂变换方法,是统计建模中常用的一种数据变换,用于连续的响应变量不满足正态分布的情况。Box-Cox变换之后,可以一定程度上减小不可观测的误差和预测变量的相关性。Box-Cox变换的主要特点是引入一个参数,通过数据本身估计该参数进而确定应采取的数据变换形式,Box-Cox变换可以明显地改善数据的正态性、对称性和方差相等性,对许多实际数据都是行之有效的。
ANalysis Of VAriance is a classical technique that accounts for difference of means across groups when the variance is homogeneous. An accompanying assumption is that data conform to a bell-shaped curve within each group.
ANOVA furnishes the tool needed for a quantitative evaluation of pairwise and simultaneous differences between a regional means.
这里介绍的比较简略,再补充一些。
ANOVA分析的目的是检验每个组的均值是否相同,零假设(null hypothesis)是\(H_0:\mu_1=\mu_2=...=\mu_n\)。使用ANOVA分析方法的前提是:
(各种性质的检验方法见4.1.4)
ANOVA中的一些重要概念:
单因素ANOVA步骤:
建立零假设和备择假设
\[H_0: \mu_1=\mu_2=...=\mu_c\]\(H_1: \mu_1、\mu_2...\mu_c\)不全相等
计算各组样本的均值和样本方差
\[\overline{x_j}=\frac{\sum_{i=1}^{n_j}{x_{ij}}}{n_j}\] \[s_j^2=\frac{\sum_{i=1}^{n_j}{(x_{ij}-\overline{x_j})^2}}{n_j-1}\]计算组间方差MSB
\[\overline{x}=\frac{\sum_{j=1}^{c}\sum_{i=1}^{n_j}{x_{ij}}}{\sum_{j=1}^{c}{n_j}}\] \[MSB=\frac{\sum_{j=1}^c{n_j(\overline{x_j}-\overline{x})^2}}{c-1}\]记\(\sum_{j=1}^{c}{n_j}=n_T\)
计算组内方差MSE
\[MSE=\frac{\sum_{j=1}^{c}{s_j^2}}{n_T-c}\]构造F检验量
\(F=\frac{MSB}{MSE}\)~\(F(c-1,n_T-c)\)
当F值很大时,则可以拒绝原假设。若给定显著性水平\(\alpha\),则拒绝域为\(F>F_{\alpha}(c-1,n_T-c)\)
以下两小节内容即从区域和方向两个角度检验了某数据样本各组均值的一致性。
本节主要介绍了灌溉区农场的数量(IRRF)和平均降水量之间进行回归分析的一个过程。涉及到的一些统计过程如下:
3 popular spatial weight matrices:
W:row standardized matrix C, the element \(w_{ij}=\frac{c_{ij}}{\sum{j=1}{n}{c{ij}}}\), or in matrix notation \(\bold{W}=\left[ \begin{matrix} \bold{1}^T\bold{c_1} & \cdots &0 \\ \vdots & \ddots & \cdots \\ 0 & \cdots & \bold{1}^T\bold{c_n} \end{matrix}\right]^{-1}\bold{C}\) where \(\bold{1}\) is an \(n*1\) vector of ones, \(\bold{c_j}\) is the \(j\) th row vector of C.
W即是行标准化后的矩阵C,该矩阵也可以用于计算MC和GR,在空间统计中很常见。
MCM=\((\bold{I}-\bold{1}\bold{1}^T/n)\bold{C}(\bold{I}-\bold{1}\bold{1}^T/n)\)
The pre- and post-multiplying matrix is one version of a projection matrix found throught linear statistical theory. Here it causes the first eigenfunction of C to be replaced with a constant vector (i.e., proportional to vector 1) and accompanying eigenvalue of 0, while asymptotically preserving the remaining n – 1 eigenfunctions.
Besides, spatial weights can also be based upon distance. Here are 3 common formulation:
two ways to analysis spatial heterogeneity: