16 Methodology of Empirical Econometric Modeling
- formulate real-business cycle theories with rational expectations, leading to
dynamic stochastic general equilibrium (DSGE) models as in Smets and Wouters
(2003), whereas Hildenbrand (1994, 1999) emphasizes heterogeneity of endow-
ments; Stiglitz (2003) stresses that asymmetric information can induce Keynesian
effects, and Aghionet al.(2002) argue that agents only have imperfect-knowledge
expectations. Moreover, many aspects of economic theory models can be chosen
freely, such as the units of time and forms of utility functions: indeed, Stigum
(1990) views theories as characterizing “toy agents in toy economies.” However,
when data are non-stationary, few transformations will be able to characterize the
evidence in a constant relationship. For example, linear relationships between vari-
ables, which often arise in Euler equations, seem unlikely to be good descriptions
in growing economies (see Ermini and Hendry, 2008, and Spanos, Hendry and
Reade, 2008, for tests of log versus linear dependent variables inI(1) processes).
The absence from many economic theories of some of the main sources of data
variability occurs across most research areas in economics, and although it differs
in form, is probably part of the reason for the rash of “puzzles” (i.e., anomalous
or even contradictory evidence) so beloved of the present generation of journal
editors. In microeconomics, lowR^2 values reveal that much of the variability is
not accounted for by the postulated models. That outcome is usually ascribed to
individual heterogeneity and idiosyncrasies, which can indeed generate high lev-
els of unexplained variability, but there has to be some doubt that all the major
factors have been included. In panel data studies, much observed data variation is
attributed to “individual effects,” which are removed by (e.g.) differencing or devi-
ations from individual means. However, if the evidence that most micro-variability
is due to individual heterogeneity is correct, then “representative” agent theories
cannot be the best basis for macro-behavior, although aggregation could sustain
some approaches (see, e.g., Granger, 1987; Blundell and Stoker, 2005), but not
others (Granger, 1980). Finally, cross-country studies rarely account for key insti-
tutional differences between the constituent economies, and often use averages
of data over historical epochs where considerable changes occurred between peri-
ods (see, e.g., Sala-i-Martin, 1997, and the criticisms in Hoover and Perez, 2004;
Hendry and Krolzig, 2004).
1.4.1.1 Non-stationarity andceteris paribus
“You don’t know how to manage Looking-glass cakes,” the Unicorn
remarked. “Hand it round first, and cut it afterwards.” (Lewis Carroll,
1899)A time series process is non-stationary if its moments or distributional form change
over time. Two important forms of non-stationarity are unit roots and structural
breaks, both of which lead to permanent changes. The former induce stochastic
trends, which can be eliminated by differencing, or cointegration can also remove
unit roots and retain linear combinations of levels of the variables (however, unit
roots and cointegration are only invariant under linear transformations of vari-
ables). There is a vast literature, and recent surveys include Hendry and Juselius