David F. Hendry 13- that all entities treated as parameters are constant over time, and invariant to
all potentially omitted variables and regime changes - the errors have “independent,” homoskedastic, distributions
- all expectations formulations are correct, or agents’ expectations are accurately
measured - the choice of estimator is appropriate at relevant sample sizes
- a valid and non-distortionary method of model selection is used.
If all of these assumptions had to be perfectly correct to produce useful empirical
evidence, there would be no hope of ever doing so. In Hendry (1987), I suggested
the four “golden prescriptions” of econometrics, abbreviated here as:
(i) think brilliantly: if you think of the right answer before modeling, then the
empirical results will be optimal and, of course, confirm your brilliance;
(ii) be infinitely creative: if you do not think of the correct model before
commencing, the next best is to think of it as you proceed;
(iii) be outstandingly lucky: if you do not think of the “true model” before starting
nor discover it en route, then luckily stumbling over it before completing the
study is the final sufficient condition. This may be the most practical of these
suggestions. Failing this last prescription:
(iv) stick to doing theory.
Lest the reader thinks the list of a dozen requirements above is overly dramatic, or
even new, Hendry and Morgan (1995) record:
In the thesis as a whole, Koopmans (1937) assembles together and confronts
most of the major issues in econometrics, which we have translated into current
terminology as:- the joint occurrence of errors-in-variables and errors-in-equations
- the need for a complete set of determining variables to leave an innovation
error - a reductionist approach of proceeding from general to simple
- the distinctions between the activities of specification, estimation and
distribution, as spelt out by R.A. Fisher - the non-experimental nature of economic data
- the need to condition on systematic components with independently
varying error terms - the choice of functional form, using linearity for convenience
- the formulation of the parameters of interest
- the need to test underlying assumptions
- the importance of incorporating all relevant information
- the avoidance of unnecessary assumptions
- the need for the general model to be estimable
- the need for the model specification to be robust.