Abstract
For testing the location parameter $\mu$ of the two parameter exponential model, a broad class of uniformly most powerful unbiased tests for the hypothesis of $\mu=0$ versus $\mu<0$ is derived. Examples are also given to show that for testing this model, there is no uniformly most powerful test for the hypothesis of $\mu=0$ versus either $\mu \neq 0$ or $\mu<0$ as claimed in the literature. Received: December 26, 2023; Accepted: February 24, 2024.
Introduction
In the field of statistical hypothesis testing, the search for a Uniformly Most Powerful (UMP) test represents the pursuit of the "best" possible procedure—one that maintains the highest probability of correctly rejecting a false null hypothesis across all alternative scenarios. For the two-parameter exponential distribution, a model frequently employed in reliability engineering and survival analysis, several foundational textbooks and papers (including the influential works of E.L. Lehmann) have long claimed the existence of UMP tests for the location parameter $\mu$.
However, recent analysis by Wang and Li identifies a significant discrepancy in these long-accepted results. This paper demonstrates that the tests previously labeled as UMP in the literature are, in fact, only Uniformly Most Powerful Unbiased (UMPU) tests. Furthermore, the authors prove that for certain hypotheses regarding the location parameter, no UMP test exists at all.
The "Conditional" Trap and Theoretical Framework
The historical error identified by the authors stems from a subtle logical leap often made when dealing with non-regular statistical models. In many standard distributions, a test that is UMP conditional on a sufficient statistic is also UMP globally. Previous literature (e.g., Takeuchi, Lehmann) applied this logic to the exponential model. Wang and Li argue that for this specific distribution—where the range of the data is dependent on the parameter—conditional optimality does not automatically translate to global optimality.
The authors employ a rigorous theoretical approach, deriving exact power functions for the standard tests using order statistics ($X_{(1)}, X_{(2)}, \dots, X_{(n)}$) and indicator functions. By comparing these functions against alternative constructions, they identify scenarios where "biased" tests—those that might have a lower probability of rejecting the null when the alternative is true than when the null is true—actually outperform the "best" unbiased tests at specific points in the parameter space.
Disproving the Existence of UMP Tests
To disprove the UMP property, the authors construct specific numerical counterexamples. For a two-sided hypothesis ($H_0: \mu = 0$ vs. $H_1: \mu \neq 0$), they consider a scenario with a sample size $n=2$ and a significance level $\alpha = 0.25$.
At the specific parameter coordinate $(\mu=1, \lambda=10)$, the "standard" test ($\phi_1$) cited in classical literature yields a power of approximately 0.89938. The authors propose an alternative test ($\phi_2$) which, at the same point, achieves a power of 0.9999. Because a UMP test must be superior to all other possible tests across the entire parameter space, the existence of $\phi_2$ definitively proves that the standard test $\phi_1$ is not UMP. The authors apply similar logic to one-sided hypotheses ($H_0: \mu = 0$ vs. $H_1: \mu < 0$), contradicting earlier claims by Takeuchi regarding the existence of a UMP test in that case.
A New Class of UMPU Tests
While UMP tests may not exist for these parameters, the authors successfully derive a broader, more flexible class of UMPU tests for the one-sided case. By demonstrating that these tests are "size-similar" (maintaining the correct significance level $\alpha$ across the boundary of the null hypothesis), they prove their unbiased optimality. This new class generalizes previously known results, showing that the "best" tests identified in earlier literature were actually just a specific subset of this larger class of unbiased tests.
Conclusion and Implications
This research serves as a critical correction to the statistical literature regarding non-regular models. By distinguishing between UMP and UMPU properties in the context of the exponential distribution's location parameter, the authors highlight the necessity of periodic re-examination of even the most widely accepted "standard" results. While the findings are focused on the two-parameter exponential model, they suggest that similar misconceptions may exist for other distributions where the data range is parameter-dependent, inviting further investigation into survival analysis models.
Summary updated