[SciPy-User] [OT] Bayesian vs. frequentist
Sturla Molden
sturla@molden...
Tue Feb 14 13:40:26 CST 2012
After having worked with applied statistics for ~15 years, I have
reached this conclusion... ;-)
Sturla's 20 propositions on Bayesian vs. classical statistics:
==============================================================
1. For simple data, a figure is sufficient, nobody really cares.
2. For dummy problems with known facit, Bayesian methods tend to be the
more accurate.
3. Bayesian methods include prior knowlege. A horse of 400 g is a priori
less likely than a horse of 400 kg. Frequentists say this is too subjective.
4. Bayesian methods are easier to interpret. Few understand a
frequenctist confidence interval, albeit everybody they think they do.
5. Hypothesis testing: Bayesians answer the question we ask. Freuentists
don't.
6. Economists investing their own money are bayesians.
7. Economists investing your money are frequentists.
8. For basic medical research, nobody cares.
9. Drug trials: For getting an FDA application approved, frequentists
often yield a more 'significant result'.
10. Drug trials: For in-hose liability estimates, Bayesian methods are
the safer.
11. Frequentists can always get more significant results by "sampling
more data".
12. Frequentists don't care about stopping rules, even though they should.
13. Bayesians don't care about stopping rules bacause they don't have to.
14. "Significant" does not mean "important". Any tiny difference can be
made statistically significant.
15. For interpreting clinical lab tests, Bayesian methods prevail, e.g.
predictive values.
16. Engineers who know their mathematics use Bayesian methods.
17. Social scientists who don't know their mathematics are frequentists.
18. SPSS, Excel, Minitab, and SAS make it easy to be an ignorant
frequentist.
19. No tool make it easy to be an ignorant bayesian.
20. Competent analysts use R, Fortran, Matlab or Python.
More information about the SciPy-User
mailing list