I'm reading Casella-Berger chapter 10, where they introduce asymptotic evaluations. I don't seem quite to understand how the factor $\sqrt{n}$ works when we are using asymptotic evaluations in order to estimate the approximate variance of estimators.

Following their example (on p. 473, here's a screenshot), given that, according to the asymptotic efficiency of MLEs (where $\hat{\theta}$ is a MLE from an iid sample, $\tau(.)$ is continuous and $I(\theta)$ is the information number for a single observation):

$\sqrt{n} \left[ \tau(\hat{\theta}) - \tau(\theta) \right] \rightarrow_d N \left(0, \frac{1}{I(\theta)}\right)$

and that, according to the Delta method, if $\sqrt{n} \left[ \hat{\theta} - \theta \right] \rightarrow_d N\left(0, \sigma^2\right)$, then:

$\sqrt{n} \left[ h(\hat{\theta}) - h(\theta) \right]\rightarrow_d N(0, \sigma^2[h^{'}(\theta)]^2)$

Why is it that (according to Casella-Berger, p. 473), the approximate variance of the estimator is given by:

$Var(h(\hat{\theta})|\theta) \approx \frac{[h^{'}(\theta)]^2}{I_n(\theta)}$

instead of:

$Var(h(\hat{\theta})|\theta) \approx \frac{[h^{'}(\theta)]^2}{I_n(\theta)} \cdot \frac{1}{n}$

I can't understand why, in this example and others (e.g. at p.243, example 5.5.25), the approximate variance of the estimator doesn't get multiplied by the $\frac{1}{n}$ term. Can anyone help me understand? Thanks in advance.

  • 2
    $\begingroup$ You are confusing the unit-level and sample-level information. The first expression is you write $\sqrt n\{\tau(\widehat \theta) - \tau(\theta)\} \to N(0,I_n(\theta)^{-1})$ presumes that $I_n$ is the unit-level Fisher information, whereas your screenshot clearly has $I_n(\theta) = E \frac{\partial^2}{\partial \theta^2} L(\theta \mid \mathbf X)$ as the sample-level Fisher information, which by linearity of expectation is $n$ times the unit-level information in this case. $\endgroup$
    – guy
    Apr 29 at 15:59
  • $\begingroup$ Yep, that totally eluded me. Thank you so much, now it's clear. $\endgroup$ Apr 29 at 16:28

Briefly, $h(\cdot)$ is a change of variable, and not a statistic. In the assumption of the $\delta$-method, we already have that $\hat{\theta}$ is scaled by $\sqrt{n}$ so that the limiting distribution has a finite, non-zero variance. As a corollary we can say $\text{Var}(\theta) \approx I_n(\theta)^{-1} / n$ for large $n$, and a similar statement can be made about the change of variable, $h(\hat{\theta})$.


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.