Assumptions on test statistics to guarantee uniform asymptotic level

76 Views Asked by At

I'm trying to formalize the fundamental ideas behind test statistics in the context of hypothesis testing. In particular, I want to define different concepts of test statistics leading to different desireable qualities of tests. I'm having trouble formalizing the idea of a "uniform test statistic" yielding a test of uniformly asymptotic level (defined below) and was hoping for some input.

Setup

Let $(\mathcal{X}, \mathbb{E})$ be a measure space and let $\mathcal{P}$ be a set of probabilty measures on the space. Let furthmore $\mathcal{P}_0 \subseteq \mathcal{P}$. We imagine getting a sample of size $n$ sampled independently, i.e. some $x \in \mathcal{X}^n$ and want to determine whether any of the measures in $\mathcal{P}_0$ are acceptable distributions for the observation.

For each $n \in \mathbb{N}$, we define a test of size $n$ as a partition of $\mathcal{X}^n$ into an acceptance region $\mathcal{A}_n$ and a critical region $\mathcal{A}_n^c$. This partition is also expressed through the critical function $\psi_n: \mathcal{X}^n \to \{0,1\}$ given by

$$ \psi_n(x) = \begin{cases} 0 & \text{if $x \in \mathcal{A}_n$}\\ 1 & \text{if $x \in \mathcal{A}_n^c$} \end{cases} $$

A sequence of tests is either $(\mathcal{A}_n)_{n \in \mathbb{N}}$ or equivalently $(\psi_n)_{n \in \mathbb{N}}$. For a given level $\alpha \in (0,1)$ a sequence of tests is said to have pointwise asymptotic level if

$$ \sup_{\nu \in \mathcal{P}_0} \limsup_{n \to \infty} P_\nu(\psi_n =1) \leq \alpha $$

and uniformly asymptotic level if

$$ \limsup_{n \to \infty} \sup_{\nu \in \mathcal{P}_0} P_\nu(\psi_n =1) \leq \alpha $$

where $P_\nu$ denotes the probability assuming that $X$ has distribution $\nu$.

Pointwise asymptotic test statistic

Let $(T_n)_{n \in \mathbb{N}}$ be a sequence of functions $T_n: \mathcal{X}^n \to \mathbb{R}$. If $T_n(X^n)$ converges in distribution to the same distribution for all $\nu \in \mathcal{P_0}$, we say that $(T_n)_{n \in \mathbb{N}}$ is a pointwise asymptotic test statistic wrt. $\mathcal{P}_0$.

Letting $V$ be a random variable with the same distribution as the limit of $T_n(X)$ for any $\nu \in \mathcal{P}_0$, if $\mathcal{B}$ is a closed set such that $P(V \in \mathcal{B}) \leq \alpha$, we say that the sequence of tests

$$ \psi_n(x) = \begin{cases} 1 & \text{if $T_n(x) \in \mathcal{B}$}\\ 0 & \text{otherwise} \end{cases} $$

is based on the pointwise asymptotic test statistic $(T_n)_{n \in \mathbb{N}}$. Note that it is quite easy to show that this test has pointwise asymptotic level by an application of the Portmanteau theorem (for the second to last inequality) since

$$ \limsup_{n \to \infty} P_\nu(\psi_n = 1) = \limsup_{n \to \infty} P_\nu(T_n(X^n) \in \mathcal{B}) \leq P(V \in \mathcal{B}) \leq \alpha $$

and then taking supremums.

Uniform asymptotic test statistic

What extra assumptions are required to prove uniform asymptotic level of a test constructed as above? We would need to ensure that the $$ \sup_{\nu \in \mathcal{P}_0} P_\nu(\psi_n=1) = \sup_{\nu \in \mathcal{P}_0} P_\nu(T_n(X^n) \in \mathcal{B}) $$

are convergent to $P(V \in \mathcal{B})$. What is a sufficient condition for this to occur?

Hope someone can help!

EDIT: I'm thinking that some conditions are required on $\mathcal{P}_0$, since in the case of $\mathcal{P}$ being all distributions on $\mathbb{R}$ with finite variance and $\mathcal{P}_0$ being all distributions with mean zero, one can show that the usual $t$-test has pointwise asymptotic level but not uniform asymptotic level.