3
$\begingroup$

I am trying to understand the asymptotic distribution of the following expression under normality $$ {\hat \sigma \hat S - \sigma S} $$

Where $\sigma$ and $S$ are the population standard deviation and population skewness respectively and $\hat{\sigma}$ and $\hat{S}$ are the sample standard deviation and sample skewness respectively. Any bright ideas, references or suggestions are appreciated

I also need to understand the following covariance structure $$ {\mathop{\rm cov}} \left[ {\hat \mu - \mu ,\hat \sigma \hat S - \sigma S} \right] $$

Where $\mu$ is population mean and $\hat{\mu}$ is the sample mean

What I have is some standard results $$ \frac{{\sqrt n \left( {\hat \mu - \mu } \right)}}{\sigma }\xrightarrow{d}{\cal N}\left( {0,1} \right) $$

$$ \,\frac{{\sqrt n \left( {\hat \sigma - \sigma } \right)}}{\sigma }\xrightarrow{d}{\cal N}\left( {0,\frac{{\kappa - 1}}{4}} \right) $$

$$ \sqrt {\frac{N}{6}} \hat S\xrightarrow{d}{\cal N}\left( {0,1} \right) $$

$\endgroup$
  • $\begingroup$ Under normality you have a much stronger result than $\frac{{\sqrt n \left( {\hat \mu - \mu } \right)}}{\sigma }\xrightarrow{d}{\cal N}\left( {0,1} \right)$ -- since that is the distribution for any $n$ (i.e. you don't need to deal with asymptotic results for that). $\endgroup$ – Glen_b Oct 29 '14 at 0:40
  • $\begingroup$ By 'skewness' do you mean the moment-skewness? -- are you just estimating that by $\hat{S}=m_3/s^3$ where $m_3=\tfrac{1}{n} \sum_{i=1}^n (x_i-\overline{x})^3$? $\endgroup$ – Glen_b Oct 29 '14 at 0:54
  • $\begingroup$ That is definition of skewness I am using and a similar definition of kurtosis should you need it $\endgroup$ – Rohit Arora Oct 29 '14 at 1:06
3
$\begingroup$

I will use the established symbol for skewness, $\gamma_1$. The population Skewness is

$$\gamma_1 = \frac {\mu_3}{\sigma^3} \Rightarrow \sigma \gamma_1 = \frac {\mu_3}{\sigma^2}$$

where $\mu_3$ is the third central moment. For the normal distribution this is zero, so the product $\sigma\gamma_1$ will also be zero.

Assuming an i.i.d. normal sample of size $n$, using the sample analogues of the moments (no need to apply finite sample corrections since we are interested in the asymptotic behavior)

$$\hat \sigma \hat \gamma_1 = \frac {(1/n)\sum_{i=1}^n \left(X_i-\bar X\right)^3}{(1/n)\sum_{i=1}^n \left(X_i-\bar X\right)^2} = \frac {\hat \mu_3}{\hat \sigma^2} \xrightarrow{p} 0$$

Then (see for example Dasgupta 2008, Theorem 3.8, page 42) we have that

$$\sqrt n\hat \mu_3 \xrightarrow{d} {\cal N}(0, V_3)$$

$$V_3 = \mu_6 - \mu_3^2 -6\sigma^2\mu_4 + 9\sigma^6,\;\; \mu_k = E(X-\mu)^k $$

Using the values and relations between the central moments of the normal distribution we obtain $V_3 = 6\sigma^6$

Since moreover $\hat \sigma^2 \xrightarrow{p}\sigma^2$, we obtain

$$\sqrt n \hat \sigma \hat \gamma_1 = \frac {\sqrt n\hat \mu_3}{\hat \sigma^2} \xrightarrow{d} {\cal N}(0, 6\sigma^{2})$$

by Slutsky's lemma. Of course the above could be directly derived using again Slutsky's lemma from

$$\sqrt {\frac{n}{6}} \hat \gamma_1 \xrightarrow{d} {\cal N}\left( 0,1 \right) \Rightarrow \sqrt n \hat \sigma \hat \gamma_1 \xrightarrow{d} \sigma \sqrt 6 {\cal N}(0,1) = {\cal N}\left( {0,6\sigma^2}\right)$$

$\endgroup$
  • $\begingroup$ Just to clarify the use of Slutsky's Lemma in the last line. Is the following correct ?$$\begin{array}{l} {{\hat \sigma }^2}\xrightarrow{p} {\sigma ^2} \Rightarrow \frac{\sigma }{{\hat \sigma }}\xrightarrow{p} 1\\ \sqrt {\frac{N}{6}} {{\hat \gamma }_1}\xrightarrow{d} {\cal N}\left( {0,1} \right) \Rightarrow \frac{{\sqrt {\frac{N}{6}} {{\hat \gamma }_1}}}{{\frac{\sigma }{{\hat \sigma }}}}\xrightarrow{d}{\cal N}\left( {0,1} \right) \end{array}$$ $\endgroup$ – Rohit Arora Oct 29 '14 at 3:55
  • $\begingroup$ Also, how would I evaluate the covariance. $${\mathop{\rm cov}} \left[ {\hat \mu - \mu ,\hat \sigma \hat S - \sigma S} \right]$$. $\endgroup$ – Rohit Arora Oct 29 '14 at 4:00
  • $\begingroup$ About Slutsky's lemma: the first relation is an application of the Continuous Mapping (Mann-Wald) Theorem - not of Slutsky's (Slutsky comes into play when we have two random variables). The second is Slutsky. About the covariance: what do you mean by "evaluate"? Finite-sample? Asymptotically? Why don't you use the usual expression of covariance and see where it gets you? $\endgroup$ – Alecos Papadopoulos Oct 29 '14 at 8:19
  • $\begingroup$ It was easy to evaluate $${\mathop{\rm cov}} \left[ {\hat \mu - \mu ,\hat \sigma - \sigma } \right] = {\mathop{\rm cov}} \left[ {\hat \mu ,\hat \sigma } \right]$$ by using bivariate delta method and the joint distribution of mean and variance is well known but in this case I have no idea about any joint distributions. Hnce, can't go beyond $${\mathop{\rm cov}} \left[ {\hat \mu - \mu ,\hat \sigma \hat S - \sigma S} \right] = {\mathop{\rm cov}} \left[ {\hat \mu ,\hat \sigma \hat S} \right]$$ $\endgroup$ – Rohit Arora Oct 29 '14 at 11:33

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.