JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.. Oxford University Press and The Review of Economic Studies, Ltd. are collaborating with JSTOR to digitize, preserve and extend access to The Review of Economic Studies. This paper derives the limit distribution of the test statistics for the error components model proposed by Breusch and Pagan under the assumption of non-normal disturbances and under a sequence of local alternatives, and shows that the Breusch-Pagan tests are robust to non-normal disturbances. The paper also points out that the Breusch-Pagan tests do not make full use of the information provided by the one-sided alternative. And it proposes a one-sided alternative test for the case where time effects are absent from the model. The newly proposed test dominates the Breusch-Pagan test in the above case. (1978) showed that the asymptotic distribution of test statistics for heteroscedasticity under the normality assumption is different from that under the assumption of non-normal disturbances, and proposed alternative test statistics for non-normal disturbances. Kendall and Stuart (1979, Chapter 31) and Kendall, Stuart and Ord (1983, Chapter 37) contain other similar problems.The purpose of the present paper is along one of the above lines, and is to investigate the robustness of test statistics for the error components model to non-normal disturbances. Widely-known test statistics for this specification test are the ones recently provided by Breusch and Pagan (1980). The present paper derives the limit distribution of these test statistics by Breusch and Pagan under the assumption of non-normal disturbances and under a sequence of alternatives contiguous to the null hypothesis. It is shown that theBreusch-Pagan tests are robust to non-normal disturbances, and have an asymptotic justification without the normality assumption. That is, we can relax the normality assumption which the original authors have employed. Since the Breusch-Pagan tests are tests for covariances of disturbances rather than for variances, the above result implies that tests for covariances are robust to non-normal disturbances in this particular model. The paper also points out that the Breusch-Pagan tests do not make full use of the information provided by a one-sided alternative. And it proposes a one-sided alternative test for the case where time effects are absent from the model. The newly proposed test 681

We propose new proxy variables for monetary policy shocks in Japan for the period from July 1989 to March 2001 and investigate the effects of changes in the policy target variable on stock prices and the term structure of interest rates. We find that changes in the surprise component of the target variable significantly affect both intermediate-term and long-term interest rates. A surprise decrease in the target rate of 1% leads, on average, to a 3% increase in stock prices. The magnitudes of estimated reactions of financial variables are similar in Japan and the US. Copyright 2006 The Authors. Journal compilation Royal Economic Society 2006.

I IHTRODUCTION\\lien \re want to test equality between sets of coefficients in two h e a r regressions, tests by Chow (19GO) are the standard procedure. The tests are based on the assumption that disturbance variances of the first set of observations are equal to disturbance variances of the second set of obserrat ions. Toyoda (1974) investigated conscquences of the test by Chow under the condition of unequal variances between two sets of observations, and concluded that the test may be misleading in some situations. Uore recently, Jayatissa (1977) and Watt (1979) have proposed two alternative testing procedures, neither of which assumes equality of disturbance variances between two sets of observations. I n particular, Watt did a Monte Carlo study of the relative properties of the two tests and reported that there is a strong case for preferring his test statistic to that of Jayatissa when the numbers of both the first and the second sets of observations are 50.The present paper provides both theoretical and numerical evidence supporting the test by Watt. Following the Introduction, Section I1 points out that the Jayatissa test statistic is not unique, not efficient and relatively complicated to compute. Section 111 gives a theoretical justification for the asymptotic Wald test proposed by \Vatt. Our numerical experiments in Section I V find that the test by Watt is prefcrable to the test by Jayatissa *Manuscript received 5.9.51 ; final version rcccived 20.1.62.tThe author is grateful to Xi. Ohtani, anonymous referees and the editor of this journal for helpful comments and suggestions on an earlier version of the paper, and to Y. Inouchi for his advice on programming.