proof of the convergence of confidence intervals

1k Views Asked by At

The confidence interval can be derived intuitively by replacing the standardized mean with the standard normal and variance with sample variance, but is there a formal limit? I'm trying to prove if the following is correct:

For $z > 0$, $Z \sim N(0, 1)$, $\mathbf{P} \left( \bar{X} - z \frac{S}{\sqrt{n}} < \mu < \bar{X} + z \frac{S}{\sqrt{n}} \right) \rightarrow \mathbf{P} (- z < Z < z)$ as $n \rightarrow \infty$ (where $S$, $\bar{X}$ are the sample sd and mean for $n$ observations, $X_i$ are iid and the mean and variances exist so CLT can be applied).

So far, I've done the following:

For some RV $Y$,

$\mathbf{P} \left( Y - z \frac{S}{\sqrt{n}} < \mu \leqslant Y + z\frac{S}{\sqrt{n}} \right) =\mathbf{P} \left( - S < \frac{Y - \mu}{z /\sqrt{n}} \leqslant S \right)$. By the strong LLN $S^2 \xrightarrow{p}\sigma^2 \Rightarrow S \xrightarrow{p} \sigma$, hence $S \pm \frac{Y - \mu}{z/ \sqrt{n}} \xrightarrow{p} \sigma \pm \frac{Y - \mu}{z / \sqrt{n}}\Rightarrow S \pm \frac{Y - \mu}{z / \sqrt{n}} \xrightarrow{d} \sigma \pm\frac{Y - \mu}{z / \sqrt{n}}$, hence $\mathbf{P} \left( - S < \frac{Y - \mu}{z/ \sqrt{n}} \leqslant S \right) \rightarrow \mathbf{P} \left( - \sigma <\frac{Y - \mu}{z / \sqrt{n}} \leqslant \sigma \right)$.

But here's where I get stuck, I can't replace $Y$ with $\bar{X}$ as $\bar{X}$ varies with $n$ so there's two limits going on here. How do I proceed?

1

There are 1 best solutions below

1
On BEST ANSWER

Outline:

Inetead of $\mathbf{P} \left( - S < \frac{Y - \mu}{z/ \sqrt{n}} \leqslant S \right),$ use $\mathbf{P} \left( - z < \frac{\bar X - \mu}{S/ \sqrt{n}} = Y \leqslant z \right).$

Find the limit of $Y$ using Slutsky's Theorem (see Wikipedia, if not in your text). Use the CLT on $\bar X$ and the fact that $S$ tends to constant $\sigma$. For an approx. 95% CI with sufficiently large $n$ use, $z = 1.96$ from standard normal tables.