30 Methodology of Empirical Econometric Modeling
help mitigate problems of under-specification, it is no free lunch, as it leads to a
different set of possible problems relating to data-based model selection, discussed
in section 1.5. However, the criticism that the LDGP is too complicated for Gets to
work well must also apply to all other approaches, as they will not fare any better
in such a state of nature unless some remarkable requirements chance to hold (e.g.,
the complexity of the LDGP happens to lie in directions completely unrelated to
the aspects under study). In general, even if one simply wants to test an economic
hypothesis as to whether some effect is present, partial inference cannot be con-
ducted alone, unless one is sure about the complete absence of all contaminating
influences.
1.4.3 Data exactitude
“She can’t do Addition,” the Red Queen interrupted. “Can you do Subtrac-
tion? Take nine from eight.” “Nine from eight I can’t, you know,” Alice
replied very readily. (Lewis Carroll, 1899)No agency produces perfect data measures on every variable, and although some
observations may be both accurate and precise (e.g., specific stock market, or for-
eign exchange, transactions), most are subject to measurement errors. These can be
difficult to handle, especially when there are both revisions and changes in exacti-
tude over time, which thereby introduce an additional source of non-stationarity.
Moreover, in any given sample of time series, more recent data will be subject to
potentially larger later revisions: section 1.7.1 considers the impact of one example
of considerable data revisions.
Mapping theoretical constructs to data counterparts and measuring (or model-
ing) latent variables both raise further issues. Many commonly used macro variables
do not have established measurements, e.g., output gaps, business cycles, capacity
utilization, trade union power, etc. Even those that do, such as constructs for con-
sumption, user costs, etc., are open to doubt. These types of measurement errors are
not directly caused by inaccurate data collection, but both impinge on empirical
studies, and can change over time.
Incentives to improve data quality, coverage and accuracy were noted in section
1.3.4 (see Boumans, 2007, for recent discussions of various measurement issues).
In the absence of exact data, there must remain trade-offs between using theory to
impose restrictions on badly-measured data, using such data to reject theory spec-
ifications, or building data-based models. Again, a balance utilizing both theory
and evidence in a progressive process seems advisable.
1.4.4 Hidden dependencies
“Why, it’s a Looking-glass book of course! And if I hold it up to a glass, the
words will all go the right way again.” (Quote from Alice in Lewis Carroll,
1899)Hidden dependencies abound in all data forms, including cross-sections, time
series and panels. An important aspect of sequential conditioning in time series