SEMAINE D'ÉTUDE SUR LE ROLE DE L'ANALYSE ECONOMETRIQUE ETC. 395
different disturbance terms. Thus it is very unlikely that V(o)
will be diagonal.
Similarly, it is rather unrealistic to assume no serial cor-
relation in the disturbances. Disturbances from econometric
models do in fact tend to be serially correlated and while we
shall later argue that correlation between a given element of u,
and a different element of «, , may be small, even a diagonal
V(0) for 0>o will not help. This is especially the case if the
time lag involved in the model is small (the very situation in
which triangularity of A is relatively likely), as in such a case
the effects of a random shock due to an omitted variable are
likely to persist for more than one time period. To put it
another way, it is natural to suppose that as the time period
involved goes to zero, V(1) approaches V(o) which is certainly
not zero (1).
Moreover, there seems little direct comfort in the points
made above that it is sufficient to have B=o0 or to have both
B and all V(0) 6>0 triangular and either B or all such V(6)
with zero principal diagonals. Economy-wide models are ge-
nerally dynamic ones so that lagged endogenous variables do
appear. Further, while we shall argue below that a diagonal
V(8) for 6 >o is not quite so unreasonable as it may seem, a
triangular B matrix is wholly unlikely, since this would be a
case in which there were no feedbacks (simultaneous or lagged)
from one variable to another and economy-wide models simply
do not have such a hierarchic structure in view of the inter-
connectedness of economic activity.
It is thus evident that even if one is willing to assume a
triangular A matrix, the assumptions of the recursive model
cannot generally be taken as valid in an economy-wide econo-
metric model. This is especially true if triangularity has been
achieved by the introduction of relatively short time lags. At
the risk of over-emphasis, we repeat. Ordinary least squares
CY See GORMAR
U|
Fisher - pag. 11
