Abstract:
In fields that are mainly non-experimental, such as economics and finance,
it is inescapable to compute test statistics and confidence regions that are
motivated by previous examinations of the same data (by the researcher
himself or by the research community as a whole). We point out
the (mathematically obvious) inadequacy of the Bayesian and Neyman-Pearson
inference theories for such a practice and present elements of a general
econometric theory we call the Neoclassical approach. This approach is
immune to multiple use of the same data, due to its reliance only on the
distribution of a statistic and not on the actual data-set at hand -- modulo
approximation errors. Classical and Bayesian procedures are often numerically
equivalent to Neoclassical ones, enabling us to give new meaning to them in
light of the Neoclassical theory.