Constructing exact confidence interval for normal distribution

56 Views Asked by At

Suppose $X_1, ..., X_n, Y_1, ..., Y_m$ are independent random variables and $X_1, ..., X_n$ i.i.d.∼$\mathcal{N}(\mu_1, \sigma_1^2)$ and $Y_1, ..., Y_m$i.i.d.∼$\mathcal{N}(\mu_2, \sigma_2^2)$, with $\mu_1, \mu_2 \in \mathbb{R}$ known. For $\alpha \in (0, 1)$, how do I construct an exact $(1 − \alpha)$-confidence interval for the quantity $\frac{\sigma_1^2}{\sigma_2^2}$?

I am confused on how to construct confidence intervals, and I hope an explanation for this problem can help me on the subject. Thanks.

1

There are 1 best solutions below

1
On BEST ANSWER

In the case where the means are unknown, the usual point estimate for the variance ratio is simply $s_1^2/s_2^2$ where $$s_1^2 = \frac{1}{n-1} \sum_{i=1}^n (x_i - \bar x)^2, \quad s_2^2 = \frac{1}{m-1} \sum_{i=1}^m (y_i - \bar y)^2, \tag{1}$$ the ratio of the unbiased sample variances. Then since $$\frac{(n-1) s_1^2}{\sigma_1^2} \sim \chi_{n-1}^2 \tag{2}$$ and similarly for $Y$, the pivotal quantity $$\frac{\sigma_1^2}{\sigma_2^2} \cdot \frac{s_2^2}{s_1^2} \sim F(m-1, n-1) \tag{3}$$ is $F$-ratio distributed with $m-1$ and $n-1$ degrees of freedom, respectively. Thus the $100(1-\alpha)\%$ confidence interval is $$\left[\frac{1}{F^*_{\alpha/2}(n-1,m-1)} \frac{s_1^2}{s_2^2}, \frac{1}{F^*_{\alpha/2}(m-1,n-1)} \frac{s_1^2}{s_2^2}\right] \tag{4}$$ where $F^*_{\alpha/2}$ is the upper $\alpha/2$ quantile of the $F$-ratio distribution, namely $\Pr[F > F^*_{\alpha/2}] = \alpha/2$.

However, when $\mu_1, \mu_2$ are known, then the estimators $(1)$ are not efficient, and the resulting interval $(4)$ may have a larger than nominal coverage probability. Instead, one may consider the alternative unbiased estimators $$\varsigma_1^2 = \frac{1}{n} \sum_{i=1}^n (x_i - \mu_1)^2, \quad \varsigma_2^2 = \frac{1}{m} \sum_{i=1}^m (y_i - \mu_2)^2. \tag{5}$$ In such a case, I believe (but have not bothered to prove it) that the pivotal quantity in $(3)$ becomes $$\frac{\sigma_1^2}{\sigma_2^2} \cdot \frac{\varsigma_2^2}{\varsigma_1^2} \sim F(m,n)$$ and the corresponding interval estimate becomes $$\left[\frac{1}{F^*_{\alpha/2}(n,m)} \frac{\varsigma_1^2}{\varsigma_2^2}, \frac{1}{F^*_{\alpha/2}(m,n)} \frac{\varsigma_1^2}{\varsigma_2^2}\right]. \tag{6}$$

At the moment I do not have the time to verify this result through simulation but it should be quite straightforward, so I will leave such details to the reader.