THE SMALL-SAMPLE PERFORMANCE OF THE INFORMATION-MATRIX TEST

This paper reviews the Information-Matrix (IM) test procedure [see White (1982)]. It investigates the size performance of several variants of the IM test statistic and, in particular, exposes the 'nR2' variant of the test, as proposed by Chesher (1983) and Lancaster (1984), as being extremely poor. Alternative 'nR2' IM test statistics are considered and it appears that substantial improvements can be made by incorporating expected values of third derivatives of the log-density into the IM test calculation procedure. Interestingly, these are exactly the sort of calculations that the Chesher (1983) and Lancaster (1984) variant avoids. © 1990.