Author: Eiko

Time: 2025-01-03 11:35:04 - 2025-01-03 12:34:16 (UTC)

Reference:

  • Theoretical Statistics by Robert W. Keener

  • Probability Theory by Eiko

  • Foundations of Modern Probability by Olav Kallenberg

Uniformly Most Powerful Tests

  • A test φ with level αφ=α is called uniformly most powerful if for any θΩ1, we have

    μθ(φ)supαφαμθ(φ)θΩ1.

    This means we are not just optimizing for one θ1 nor for the power function, we are optimizing for all θ1Ω1. It might not exist.

  • A family of densities pθ(x), θΩR has monotone increasing likelihood ratios if there exists a statistic T=T(x) such that whenever θ1<θ2, the likelihood ratio pθ2(x)/pθ1(x) is a non-decreasing function of T(x).

    For the decreasing case it suffices to replace T by T.

Example. For densities of the form of exponential families (of single parameter)

μθexp(η(θ)T(x)B(θ))h(x)dx,

if η(θ) is non-decreasing / increasing, the likelihood ratio

L(x)=dμθ2dμθ1=exp((η(θ2)η(θ1))T(x)B(θ2)+B(θ1))

is a non-decreasing / increasing function of T(x). This includes the example of the exponential distribution, for which η(θ)=θ is increasing and T(x)=x.

Monotone Likelihood Gives Uniformly Most Powerful Tests

For this concept of monotone likelihood, we assume the parameter space θΩR is one-dimensional.

Suppose the family of density P(X|θ) has monotone likelihood ratio (with respect to a test statistic T(x)). Then

  • The monotone T test

    φ(x)={1,T>cγ,T=c0,T<c

    is a uniformly most powerful test for the hypothesis H0:θθ0 and H1:θ>θ0, with level

    α=μθ0φ.

    c and γ can be adjusted to achieve any desired level α.

  • The power function β(θ)=μθ(φ) for this test is strictly increasing when it is in (0,1), and non-decreasing overall.