On likelihood ratio tests

Likelihood ratio tests are intuitively appealing. Nevertheless, a number of examples are known in which they perform very poorly. The present paper discusses a large class of situations in which this is the case, and analyzes just how intuition misleads us; it also presents an alternative approach which in these situations is optimal. 1. The popularity of likelihood ratio tests Faced with a new testing problem, the most common approach is the likelihood ratio (LR) test. Introduced by Neyman and Pearson in 1928, it compares the maximum likelihood under the alternatives with that under the hypothesis. It owes its popularity to a number of facts. (i) It is intuitively appealing. The likelihood of θ, Lx(θ) = pθ(x) i.e. the probability density (or probability) of x considered as a function of θ, is widely considered a (relative) measure of support that the observation x gives to the parameter θ. (See for example Royall [8]). Then the likelihood ratio sup alt [pθ(x)]/ sup hyp [pθ(x)] (1.1) compares the best explanation the data provide for the alternatives with the best explanations for the hypothesis. This seems quite persuasive. (ii) In many standard problems, the LR test agrees with tests obtained from other principles (for example it is UMP unbiased or UMP invariant). Generally it seems to lead to satisfactory tests. However, counter-examples are also known in which the test is quite unsatisfactory; see for example Perlman and Wu [7] and Menéndez, Rueda, and Salvador [6]. (iii) The LR test, under suitable conditons, has good asymptotic properties. None of these three reasons are convincing. (iii) tells us little about small samples. (i) has no strong logical grounding. (ii) is the most persuasive, but in these standard problems (in which there typically exist a complete set of sufficient statistics) all principles typically lead to tests that are the same or differ only by little. Department of Statistics, 367 Evans Hall, University of California, Berkeley, CA 94720-3860, e-mail: [email protected] AMS 2000 subject classifications: 62F03.