SEMAINE D'ÉTUDE SUR LE ROLE DE L’ANALYSE ECONOMETRIQUE ETC. 183
sumptions of interdependent systems, this requirement will actually
be fulfilled (1).
I welcome very much Professor KooPMANS’ questions about the
notion of autonomy. Prof. FRISCH’s original concept of autonomy
refers to an economic feature, for example a parameter B, that re-
mains invariant when other things change; this concept is closely
related to the notion of invariance as formulated in Professor Haa-
VELMO's previous comment, Professor KooPMANS refers to auto-
nomy in a related sense that emphasizes-the model aspects, a relation
being called autonomous if it can be broken out of a model and
inserted in some other specified model. In such autonomous rela-
tions the parameters could be called autonomous in the sense of
the above points (2) and (3). As regards the argument about a
change in policy, the autonomy refers to the case when the model
builder uses different models before and after the change.
In reply to Professor Koopmans, it is my understanding that
the notion of autonomous relations is highly relevant for the theory
of multirelation models in general, and in particular so for ID, CC
and BEID systems. The difference in approach may perhaps call
for some slight modification, mutatis mutandis, depending upon what
type of model we are considering. Thus for ID systems, the auto-
nomy of a behavioural relation would require that it remains the
same if it is broken out and inserted in another system which inclu-
des those current endogenous variables that enter as explanatory
in the autonomous relation. For CC and BEID systems the concept
of autonomy might well be generalized somewhat so as to require
only that the residual-free part of the relation remains the same when
it is inserted in some other model. This Jast remark emphasizes the
point I wish to make in (2), namely that the invariants of primary
importance in non-experimental model building are directed predic-
tors such as (E) rather than joint probability laws.
('} This is however not the whole story, for it turns out that the
classic assumptions constitute a special case that covers only a subspace
of lower dimension than the entire parameter space: see reference (b) in
footnote
Wold
pag. 69