On the Proportion of Time that a Conventional Confidence Interval Covers

Abstract

The concept of confidence sequences, or `anytime confidence intervals,' has attracted significant interest in recent years. This motivates the question of how well conventional confidence intervals behave when similarly evaluated. In particular, we study the probability distribution of the proportion of time a conventional confidence interval procedure covers the population mean, when the total time $n$ spent sampling is large. We establish an invariance principle for the limiting distribution as $n$ tends to infinity and show that the limit random variable $Z(1)$ can be expressed as a functional of Brownian motion. The random variable $Z(1)$ exhibits a scale invariance property that enables the derivation of a partial differential equation (PDE) for the density of $Z(1)$ jointly with the Brownian motion at $t=1$, namely $B(1)$. We show that this density can also be computed as the stationary distribution of a two-dimensional singular diffusion, one component of which is an Ornstein-Uhlenbeck process. We then derive an approximating Markov chain that allows us to numerically solve the PDE. This density has a number of interesting features that provide various insights into the time-averaged coverage proportion problem.