Tests for parameter changes at unknown times in linear regression models (Q1174646)

The problem of testing for parameter changes at unknown time points in linear regression models is considered. The approach taken for testing a null hypothesis of no changes against one- as well as two-sided alternative hypotheses is based on a `Bayes-type' statistic. In the suggested approach an appropriate prior distribution on nuisance parameters associated with both the null and the alternative hypotheses is first assumed. Then, the respective unconditional likelihood functions are obtained through elimination of the nuisance parameters, and the likelihood ratio statistic is derived. The Bayes-type statistics for one- and two-sided alternatives, and their asymptotic distributions are derived. In addition, a simulation study is presented ``with the aim of comparing powers of the Bayesian-type statistics with other statistics in the literature''. The paper also includes an application of the test procedure to data on AIDS in the United States. It is here important to note that in Bayesian inference, the parameters are regarded as random variables with different probability distributions, or degrees of beliefs. Therefore, the test proposed is desirable if one wishes to test whether his beliefs about parameter changes in a linear regression model have changed (i.e. testing for self consistency). This should not be mixed up with the frequentist view of inference, in which the parameters have fixed but unknown values, and are tested by using test statistics. The probability distribution of a test statistic, the sampling distribution, is interpretable only in relation to repeated realizations of the current situation. My concern about this paper is therefore that the authors do not make the impossibility of a frequentist interpretation of the proposed test clear enough. Readers of the paper might erroneously believe that conclusions drawn from the tests have an objective interpretation.