Rp. Byron, DIAGNOSTIC TESTING AND SENSITIVITY ANALYSIS IN THE CONSTRUCTION OF SOCIAL ACCOUNTING MATRICES, Journal of the Royal Statistical Society. Series A. Statistics in society, 159, 1996, pp. 133-148
This paper examines the issue of testing for initial estimate bias in
the construction of a social accounting matrix (SAM). The issue arises
because the statistician may have inadvertently provided incorrect in
itial estimates through simple human error, underreporting, miscategor
ization or for any of a host of possible reasons. Baxter has made a st
art on the subject, using only the Mahalanobis distance (or Wald test)
as the basis for inference. The tests available fall into the standar
d likelihood ratio-Lagrange multiplier-Wald categorization and, as exp
ected, display good power in identifying a biased cell estimate. Howev
er, the problem is much more complicated than raised by Baxter and the
present paper only addresses some of the complications. How can tests
be used to identify biased initial estimates? What happens to the tes
ts as the size of an SAM increases? Which of the three tests is to be
preferred? The simplest procedure, that of comparing the balanced with
the unbalanced initial estimate within the context of the variance as
signed to the initial estimate, is shown to be a likelihood ratio test
. The performance of the tests does not appear to diminish as the size
of the SAM increases, probably because the number of random terms int
roduced increases at a faster rate than the number of restrictions (th
e size of the SAM). The Wald and Lagrange multiplier tests of a cell r
equire a joint test of a row and column restriction simultaneously; ho
wever, Monte Carlo experiments suggest the counter-intuitive result th
at the difference (likelihood ratio) test based on the restricted and
unrestricted estimate of a cell may be superior to either. The methods
developed here have relevance to other areas of data construction, su
ch as national accounting or the reconciliation of international trade
statistics.