188
where
n-k
1 atat-k
k-1
Z a?
k-1
where n is the number of observations in the series, k is the lag
length, m is the maximum lag, and at is the residual series. For vir
tually every series for which the null hypothesis was rejected, first
order autocorrelation was evident. Thus, the polynomial A(B) defined in
Chapter V is not simply equal to one. A number of methods are available
that correct for serial correlation. This study utilized both the
Cochran-Orcutt method and addition of a lagged dependent variable. The
inclusion of a dependent variable lagged one period, however, gave more
promising results, based on the resulting Ljung-Box statistic. The
results of the procedure for the monthly and quarterly data for both
size classes are given in Table B. Therefore, an additional lagged
endogenous price may occur in the of each price model due to statis
tical considerations.
The efficient estimation of systems of equations (this study
employed recursive and simultaneous systems) requires that the cross
correlation of error terms must be taken account. In particular, the
contemporaneous cross correlation among error texrms in a system of
models must theoretically be zero. Mehta and Swamy (1976), however,
suggest a less rigid criterion, e.g. that the contemporaneous cross
correlation between error terms be less than or equal to .30. When this
is not the case, single equation methods and 2SLS are no longer effi
cient. Instead, single equation and 2SLS methods should be replaced by
seemingly unrelated estimation and 3SLS methods, respectively (Pindyck