Find the confidence interval of square of the mean $\mu^2$

564 Views Asked by At

Suppose I have a population, I know the variance $\sigma^2$, but I don't know the mean $\mu$. How to find the confidence interval for the square of the mean $\mu^2$ ?

It's well known that the CI of the mean $\mu$ is $P(a < \mu < b) = 1-\alpha$ where $a = \bar x- z_{\alpha/2}\cdot\frac{\sigma}{\sqrt{n}}$ and $b = \bar x+ z_{\alpha/2}\cdot\frac{\sigma}{\sqrt{n}}$.

Since $P(a < \mu < b)=P(2a < 2\mu < 2b) = 1-\alpha$, can I just use $P(a^2 < \mu^2 < b^2) = 1-\alpha$ to find the CI of $\mu^2$?

Similarly, if I want to find the CI of $\mu^2+2\mu$, is it correct to do this: $$P(a^2+2a < \mu^2+2\mu < b^2+2b) = 1-\alpha\,\,?$$

And also, what if I'm not sure that the population follows the standard normal distribution? Is there any generalized formula for the CI?

1

There are 1 best solutions below

0
On BEST ANSWER

By CLT,

$$ \sqrt{n}\frac {\bar{X} - \mu} {\sigma} \to \mathcal{N}(0, 1) $$

Here the key is you need to recognize that this is a (asymtotic) pivotal quantity: the distribution is independent of the unknown parameter $\mu$, and thus we can compute the quantile $z_{\alpha/2}$, and then invert it to construct the CI.

To construct a similar pivotal quantity involving $\mu^2$, we may consider the sample $X_i^2$, with

$$ E[X^2] = Var[X] + E[X]^2 = \sigma^2 + \mu^2$$

Since $Var[X^2]$ involve the fourth moment, we just use the sample variance $S_{X^2}^2$ to estimate it, as it is the consistent estimator of $Var[X^2]$:

$$ S_{X^2}^2 = \frac {1} {n - 1} \sum_{i=1}^n \left(X_i^2 - \bar{X^2}\right)^2$$

where

$$ \bar{X^2} = \frac {1} {n} \sum_{i=1}^n X_i^2 $$

is the sample mean. Then by Slutsky and CLT,

$$ \sqrt{n} \frac {\bar{X^2} - \sigma^2 - \mu^2} {S_{X^2}} \to \mathcal{N}(0, 1)$$

Hence,

$$ \begin{align} \Pr\left\{-z_{\alpha/2} \leq \sqrt{n} \frac {\bar{X^2} - \sigma^2 - \mu^2} {S_{X^2}} \leq z_{\alpha/2} \right\} \to 1 - \alpha \\ \Rightarrow \Pr\left\{-z_{\alpha/2} \frac {S_{X^2}} {\sqrt{n}} \leq \bar{X^2} - \sigma^2 - \mu^2 \leq z_{\alpha/2} \frac {S_{X^2}} {\sqrt{n}} \right\} \to 1 - \alpha \\ \Rightarrow \Pr\left\{\bar{X^2} - \sigma^2 -z_{\alpha/2} \frac {S_{X^2}} {\sqrt{n}} \leq \mu^2 \leq \bar{X^2} - \sigma^2 + z_{\alpha/2} \frac {S_{X^2}} {\sqrt{n}} \right\} \to 1 - \alpha \\ \end{align} $$

As a result, $$ \left[\bar{X^2} - \sigma^2 - z_{\alpha/2} \frac {S_{X^2}} {\sqrt{n}}, \bar{X^2} - \sigma^2 + z_{\alpha/2} \frac {S_{X^2}} {\sqrt{n}} \right] $$

is a asymptotic $(1 - \alpha)$ confidence interval for $\mu^2$.

Note that it is just one way to construct the CI; there are other possibilities as well.