How to quantify estimation uncertainty and propagate it through transformations
A point estimate gives one number; a confidence interval gives a range that communicates uncertainty. The Delta method lets us derive confidence intervals for transformed parameters when we know the asymptotic distribution of the original estimator.
A confidence interval \(\mathcal{I}\) of level \(1-\alpha\) for \(\theta\) is a random interval such that:
\[\mathbb{P}_\theta(\theta \in \mathcal{I}) \geq 1-\alpha \quad \forall\,\theta \in \Theta\]
An asymptotic CI of level \(1-\alpha\) satisfies \(\lim_{n\to\infty}\mathbb{P}_\theta(\theta \in \mathcal{I}) \geq 1-\alpha\).
If \(\sqrt{n}(\hat{\theta}_n - \theta^*) \xrightarrow{(d)} \mathcal{N}(0,\sigma^2)\), then an asymptotic \(1-\alpha\) CI is:
\[\mathcal{I} = \left[\hat{\theta}_n - q_{\alpha/2}\frac{\hat{\sigma}}{\sqrt{n}},\;\; \hat{\theta}_n + q_{\alpha/2}\frac{\hat{\sigma}}{\sqrt{n}}\right]\]where \(q_{\alpha/2}\) is the \((1-\alpha/2)\)-quantile of \(\mathcal{N}(0,1)\) and \(\hat{\sigma}\) is a consistent estimator of \(\sigma\).
Observe \(R_1,\ldots,R_n \stackrel{iid}{\sim} \text{Ber}(p)\). By CLT: \(\sqrt{n}(\bar{R}_n - p)/\sqrt{p(1-p)} \xrightarrow{(d)} \mathcal{N}(0,1)\). The 95% CI for \(p\) uses \(\hat{p} = \bar{R}_n\):
\[\mathcal{I} = \left[\hat{p} - 1.96\sqrt{\frac{\hat{p}(1-\hat{p})}{n}},\;\; \hat{p} + 1.96\sqrt{\frac{\hat{p}(1-\hat{p})}{n}}\right]\]The Delta method propagates asymptotic normality through smooth transformations. If we know the distribution of \(\hat{\theta}_n\), and we want the distribution of \(g(\hat{\theta}_n)\) for some smooth function \(g\):
Let \((Z_n)_{n\geq1}\) be a sequence such that \(\sqrt{n}(Z_n - \theta) \xrightarrow{(d)} \mathcal{N}(0,\sigma^2)\). Let \(g: \mathbb{R}\to\mathbb{R}\) be continuously differentiable at \(\theta\). Then:
\[\sqrt{n}(g(Z_n) - g(\theta)) \xrightarrow{(d)} \mathcal{N}(0,\, (g'(\theta))^2\sigma^2)\]
Intuition: Taylor-expand \(g\) around \(\theta\): \(g(Z_n) \approx g(\theta) + g'(\theta)(Z_n - \theta)\). Since the linear term dominates, the asymptotic variance scales by \((g'(\theta))^2\).
From \(\bar{X}_n \approx \mathcal{N}(\mu, \sigma^2/n)\), what is the distribution of \(\bar{X}_n^2\)?
Apply \(g(x) = x^2\), so \(g'(x) = 2x\), \(g'(\mu) = 2\mu\). By the Delta method:
\[\sqrt{n}(\bar{X}_n^2 - \mu^2) \xrightarrow{(d)} \mathcal{N}(0, 4\mu^2\sigma^2)\]So the asymptotic variance of \(\bar{X}_n^2\) is \(4\mu^2\sigma^2/n\).
Let \((T_n)\) be a sequence of random vectors such that \(\sqrt{n}(T_n - \theta) \xrightarrow{(d)} \mathcal{N}_d(0,\Sigma)\). Let \(g: \mathbb{R}^d \to \mathbb{R}^k\) be continuously differentiable. Then:
\[\sqrt{n}(g(T_n) - g(\theta)) \xrightarrow{(d)} \mathcal{N}_k\!\left(0,\, \nabla g(\theta)^\top \Sigma\, \nabla g(\theta)\right)\]
The gradient \(\nabla g(\theta)\) replaces \(g'(\theta)\), and the covariance transforms by the Jacobian. This is used to derive CIs for functions of multiple parameters, like ratios or products.