Most of the statistics textbooks claim that point estimator $\hat\theta_n(\mathbf{X})$ is asymptotically normal if $$\sqrt{n}(\hat\theta_n - \theta) \, \overset{d}{\longrightarrow} \, \mathcal{N}(0, \sigma^2(\theta)), \qquad \forall \theta \in \Theta$$ where $\sigma^2(\theta)$ is called asymptotic variance of $\hat\theta_n$.

Moreover, if $\sigma(\theta)$ is a continuous function of $\theta$ ($\forall \theta \in \Theta$), we can write that

$$\frac{\sqrt{n}(\hat\theta_n - \theta)}{\sigma(\theta)} \, \overset{d}{\longrightarrow} \, \mathcal{N}(0, 1), \qquad \forall \theta \in \Theta$$

But Larry Wasserman in his book "All of statistics" give the following definition of the asymptotically normal estimator:

An estimator is asymptotically Normal if $$\frac{\hat\theta_n - \theta}{\mathrm{SE}} \, \overset{d}{\longrightarrow} \, \mathcal{N}(0, 1)$$

So we have that $\mathrm{SE} = \frac{\sigma(\theta)}{\sqrt{n}}$.

Also Wasserman previously defined standard error as $\mathrm{SE} = \sqrt{\mathrm{Var}_\theta(\hat\theta)}$.

Therefore, for any asymptotically normal estimator $\hat\theta_n$ I concluded that $$\mathrm{SE} = \sqrt{\mathrm{Var}_\theta(\hat\theta)} = \frac{\sigma(\theta)}{\sqrt{n}} \qquad (*)$$

I know that $(*)$ is true when $\hat\theta_n$ is a sample mean (in that case $\sigma(\theta) \equiv \sigma = \sqrt{\mathrm{Var}(X)}$)

But is $(*)$ true for any asymptotically normal estimator ?

P.s. In some sources I have seen that quantity $\frac{\sigma(\theta)}{\sqrt{n}}$ is called "asymptotic standard error" and actively used in building confidence intervals. Also please keep in mind that I am not talking about the estimated standard error $\widehat{\mathrm{SE}}(\hat\theta_n) = \sqrt{\widehat{\mathrm{Var}}(\hat\theta_n)}$ which is an observed value.

EDIT. mpiktas's answer seems to be the truth.
Indeed, the equality $(*)$ is correct only for sample mean $\overline{X}$. If we get another asymptotically normal estimator, for example sample variance $S^2 =\frac{1}{n-1} \sum_{i=1}^n (X_i - \overline{X})^2$, then the equality $(*)$ will not hold. It is easy to show.
On the one hand, $$\mathrm{SE}(S^2) = \sqrt{\mathrm{Var}(S^2)} = \sqrt{\frac{1}{n} \left(\mu_4 - \frac{n-3}{n-1}\sigma^4\right)}, \qquad (1)$$ where $\mu_4 = \mathrm{E}[(X_1 - \mu)^4]$ (see wiki, for example).

On the other hand, $$\frac{\sigma(S^2)}{\sqrt{n}} = \frac{\sqrt{\mu_4 - \sigma^4}}{\sqrt{n}}, \qquad (2)$$ where $\sigma(S^2)$ is a square root of the asymptotic variance of $S^2$ (see proof here).

Expressions (1) and (2) are not equal, i.e. $\mathrm{SE}(S^2) =\sqrt{\mathrm{Var}(S^2) } \neq \sigma(S^2)/\sqrt{n}$.

But we have the following convergence: $$\sqrt{n} \, \mathrm{SE}(S^2) \, \overset{n \to \infty}{\longrightarrow} \, \sigma(S^2) = \sqrt{\mu_4 - \sigma^4}.$$


The answer is no.

Your definition is wrong. SE is calculated from data, $\sigma(\theta)$ is not observed and is a theoretical concept. So there cannot be a equality.

The correct conclusion which you should have made is the following

$$\sqrt{n}SE \to \sigma$$

for the asymptotically normal estimator $\hat\theta$. This is correct for all asymptotically normal estimators. (Note there might some weird cases when this does not hold related to intricacies of CLT, but since I cannot recall them at the moment I am putting this disclaimer just in case).

| cite | improve this answer | |
  • $\begingroup$ Thanks for the answer. I want to clarify - correct definition of SE is the following: $\mathrm{SE}(\hat\theta_n) = \sqrt{\mathrm{Var}(\hat\theta_n)}$, and generally $\sqrt{\mathrm{Var}(\hat\theta_n)} \neq \frac{\sigma(\theta)}{\sqrt{n}}$, right? (we can only say that $\sqrt{n}\, \mathrm{SE}(\hat\theta) = \sqrt{n \cdot \mathrm{Var}(\hat\theta_n)} \,\longrightarrow \, \sigma(\theta)$ when $n \longrightarrow \infty$) $\endgroup$ – Rodvi Aug 31 '18 at 14:11
  • $\begingroup$ It seems that you are correct! I investigated sample variance $S^2$ to check you idea (see my edit to the post above). $\endgroup$ – Rodvi Aug 31 '18 at 20:51
  • $\begingroup$ Note that SE should be an estimate too for the definition to be useful. The whole point of asymptotic normality is to have a quantity which can be calculated which is approximately normal. $\endgroup$ – mpiktas Sep 7 '18 at 5:38

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.