Before condemning data mining one should distinguish between objective and biased data mining. The former is commendable. Even biased data mining is appropriate when used to illustrate and not to test hypotheses. In the context of testing, the problem with biased data mining arises not from the fitting of many regression, but from inadequate reporting of results. The trend towards reporting the results of more alternative specifications, and thus addressing the fragility problem, should be encouraged. To do that the incentives that economists face should be changed.
