残值是常态分配平均值分析的假设可用QuantileRegression修正.PPT

Quantile Regression分量迴歸 1. The Problem Quantile Regression OLS的估計方法 解釋變數對被解釋變數的平均影響效果,平均影響效果之適用性 平均值正負二邊是對稱 1.1 OLS Problem A Problem A The distribution of Y, the “dependent” variable, conditional on the covariate X, may have thick tails. The conditional distribution of Y may be asymmetric. The conditional distribution of Y may be non-unimodal. Neither regression nor ANOVA will give us robust results. Outliers are problematic, the mean is pulled toward the skewed tail, multiple modes will not be revealed. OLS的估計方法缺失 很多資料型態顯然不符合常態分配。 工資 報酬率 財金資料 教育投入與薪資 大多均為 high skewness OLS的估計方法缺失 解釋變數對被解釋變數的平均影響效果。 忽視少數有時是重要的 有時少數是重要的 例如 少數的太重與太輕的出生嬰兒在生產過程是高危險群 80%的理賠來自少數的20%產險客戶 Quantile Regression Quantile Regression The Perspective of Quantile Regression (QR) 2. The Estimator Quantile Regression Discrete Distribution Definition: Given P ∈ [0, 1]. A pth quantile of a random variable Z is any number zp such that Pr(Y zp) ≤ P ≤ Pr(Y ≤ zp). The solution always exists, but need not be unique. Continuous Distribution Suppose Y is a continuous r.v. with cdf F(.), then Pr(zy) = F(y) for every y in the support and a pth quantile is any number zp such that F(zp) = p If F is continuous and strictly increasing then the inverse exists uniquely and zp= F-1(p) Quantile Regression 2.1 Central Dispersion Quantiles can be used to characterize a distribution: Median Interquartile Range Interdecile Range Symmetry = (z.75- z.5)/(z.5- z.25) Tail Weight = (z.90- ζ.10)/(z.75- z.25) 2.2 Estimation Loss Absolute Value Least Square Quantile Loss 2.2.1 Absolute Value 2.2.2 Least Square Ordinarily we specify a quadratic loss function. That is, L(e) = e2 Under quadratic loss we use the conditional mean, via regression or ANOVA, as our predictor of Y for a given X = x. 2.2.3 Quantile Loss Quantile Loss Function Lp 2. Lp = Σθe+ + Σ(1-θ)e- Loss Function Computation 1 Computation 2 2.3 Quantile Estimator Proposition: Under the asymmetric absolute loss function LP a best predictor of Y give