SEMAINE D’ETUDE SUR LE ROLE DE L'ANALYSE ECONOMETRIOUE ETC. 
176 
quent use in the debate on the rationale of interdependent systems, 
[ should like to comment a little more on this type of application. 
Nonexperimental model building cannot be based on the results of 
controlled experiments; speaking generally, the empirical basis of 
the model is instead some kind of regularity in the observed pheno- 
mena, regularities that the model builder tries to explore and explain 
by his model. In presenting his model, he should broadly specify 
these regularities as the intended domain of validity of his model. 
And if a scientific model is to be used for forecasting the results 
of a change in economic policy, the observed regularities should 
include some evidence from earlier changes in policy. There is here 
a fluid border between science and politics. Several aspects of 
science and politics have come to the fore in other sessions of out 
Study Week. The only point I wish to make in the present context 
is that politics has other social functions than science, and therefore 
political activity can never be completely rationalized as an appli- 
cation of scientific model building. 
Coming finally to my point of disagreement with Professor Haa- 
VELMO’s comments, it lies in his broad statement that a stochastic 
economic model is nothing else than a joint probability law, and 
he even goes as far as to put between quotation marks the « rela- 
tions » that can be derived as properties of the joint probability 
law that constitutes the model. True, the stochastization of de- 
terministic models is a key development in modern econometrics, 
and in this connection I was nearly to say that the part played by 
joint probability laws in the specification of nonexperimental models 
cannot be exaggerated — but the point I wish to make is just that 
Professor HAAVELMO’s statement is such an exaggeration. Joint pro- 
bability laws can express much, but they cannot express everything. 
They are symmetric in the variables involved, and as such they 
cannot express asymmetric features of the model, and in particular 
they cannot express causal relations that enter as part of the model, 
for causal relationships are directed (from cause to effect) and the- 
reby asymmetric. Causal relations in general are directed and asym 
metric both in deterministic and stochastic models, and the stochasti 
x 
Wold - pag. 63