I comment on the controversy between McCloskey & Ziliak and
Hoover & Siegler on statistical versus economic significance, in
the March 2008 issue of the Journal of Economic Methodology.
I argue that while McCloskey & Ziliak are right in emphasizing
'real error', i.e. non-sampling error that cannot be eliminated
through specification testing, they fail to acknowledge those areas
in economics, e.g. rational expectations macroeconomics and
asset pricing, where researchers clearly distinguish between statistical
and economic significance and where statistical testing
plays a relatively minor role in model evaluation. In these areas
models are treated as inherently misspecified and, consequently,
are evaluated empirically by other methods than statistical tests.
I also criticise McCloskey & Ziliak for their strong focus on the
size of parameter estimates while neglecting the important question
of how to obtain reliable estimates, and I argue that significance
tests are useful tools in those cases where a statistical
model serves as input in the quantification of an economic model.
Finally, I provide a specific example from economics - asset return
predictability - where the distinction between statistical and
economic significance is well appreciated, but which also shows
how statistical tests have contributed to our substantive economic
understanding.