Beyond HC: more sensitive tests for rare/weak alternatives (Q2215733)
From MaRDI portal
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Beyond HC: more sensitive tests for rare/weak alternatives |
scientific article |
Statements
Beyond HC: more sensitive tests for rare/weak alternatives (English)
0 references
14 December 2020
0 references
The paper under review discusses the utilization of a popular method for large-scale inference problems that is called Higher Criticism (HC). The purpose of the paper is to make the case that one should also consider the parametric methods, including the generalized likelihood ratio test (GLRT) and a related method involving the empirical moment-generating function (EMGF), as complementary methods to HC and that indeed they are closer to HC in nature than might appear at first glance. In particular, it is shown that there is another similar `narrow' model to which HC is `tied' in precisely the same way as the GLRT and EMGF are `tied' to the normal location mixture model. The authors' intention is not to discredit HC; it is rather to point out that the notion of being `tied to a narrowly specified model' is a misplaced criticism for statistics of this type. The framework under which the theoretical properties of HC were originally developed was not detailed enough to discern any difference in performance between HC and the GLRT under the normal location mixture model. The main technical contribution of this paper is to provide a framework for higher-order power comparisons between these and related statistics, which reveals that each statistic has an edge in power under the model to which it is `tied' but none is `uniformly better' across all scenarios. The main technical results (on power under sparse local alternatives) are presented under the two important examples of the contamination models. A summary of simulation experiments, used to illustrate the theoretical results, is also provided.
0 references
higher criticism (HC)
0 references
sparse normal means
0 references
phi-divergence
0 references
multiple comparisons
0 references
mixture model
0 references
0 references
0 references
0 references
0 references