The theorem is found on Stephen M. Stigler paper The Asymptotic Distribution of The Trimmed Mean (1973)

Let $S_n$ denote trimmed mean and: $$a = \sup\{x: F(x) \le \alpha\} \\ b = \inf\{x:F(x)\ge \beta\}$$ $$A = a - \inf\{x: F(x) \ge \alpha\}, \\ B = b - \sup\{x:F(x) \le \beta\}-b$$

Further define: $$\begin{matrix}G(x) = & 0 & \text{for }x<a, \\ & (F(x)-\alpha)/(\beta-\alpha) & \text{for } \alpha\le b, \\ & 1 & x\ge b\end{matrix}$$ Theorem: As $n \to \infty$, $L(n^{1/2}(S_n - \mu)) \to L(Z)$, with $Z = (\beta - \mu)^{-1} \left[Y_1 + (b-\mu) Y_2 +(\alpha - \mu) Y3 + B\max(0,Y_2) - A\max(0,Y_3)\right]$ and $E[Z] = \left[B(\beta(1-\beta))^{1/2} - A(\alpha(1-\alpha))^{1/2}\right] / \left[(\beta-\alpha)(2\pi)^{1/2}\right]$, where $Y_1\sim \mathcal{N}(0, (\beta-\alpha)\sigma^2)$ and independent of $(Y_1, Y_3)$, and $(Y2,Y3) \sim \mathcal{N}(0,C)$, where C is

$$C = \left(\begin{matrix}\beta(1-\beta) & -\alpha(1-\beta) \\ -\alpha(1-\beta) & \alpha(1-\alpha)\end{matrix}\right)$$

First, I cannot grasp what $Z$ is. In the introduction it is stated that with the help of this theorem the distribution of trimmed mean will be derived. But the theorem itself is about the asymptotic distribution of $S_n - \mu$.

Second, what is the meaning of $Y_1, Y_2 $ and $Y_3$?

Third, how could the $C$ covariance matrix be interpreted? Does the $\max(0,Y_2)$ denote the maximum variance of $Y2$ or what?

Proof. Given $X_1,...,X_n$ define random variables $V_n$ and $W_n$ as follows: if $F(a-) = \alpha$, let $V_n = (\#X_i < a);$ otherwise let $V_n$ count the number of $X_i = a$ with probability $p_1 = (\alpha - F(a-)) / (F(a) - F(a-))$ each (independently of others) plus the number of $X_i < a$. If $F(b) = \beta$, let $W_n = (\# X_i \le b)$; otherwise let $W_n$ count the number of $X_i = b$ with probability $p_2 = (\beta - F(b-)) / (F(b) - F(b-))$ each (independently of others) plus the number $X_i < b$. (In the special case $a=b$ and $F(a-) < \alpha<\beta < F(a)$, then conditional on $(\#X_i = a) = m$, perform $m$ independent trinomial trials with probabilities $(p_1, p_2 - p_1, 1-p_2)$, observing $(R_1,R_2,R_3)$. Let $V_n = (\#X_i<a)+R_1$ and $Wn = V_n + R_2$. Then $V_n$ has a binomial $(n, \alpha)$ distribution, $W_n$ has a binomial $(n, \beta)$ distribution, and $Cov(V_n, W_n) = n\alpha(1-\beta)$.

What does $V_n$ and $W_n$ signify? In the special case of $a=b$ , why is it supposed that $F(-a) < F(a)$? Also, regarding the special case where $a=b$, it seems to me that $\alpha$ must be equal to $\beta$, because $a=b \Rightarrow \sup\{x: F(x) \le \alpha\} = \inf\{x:F(x)\ge \beta\}$, and if $\alpha \neq \beta$, then the distribution is with two blocks of data on the sides and then one point in the middle, isn't it?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.