[SciPy-User] [OT] Bayesian vs. frequentist
Wed Feb 15 03:00:12 CST 2012
On 02/15/2012 09:21 AM, Daniele Nicolodi wrote:
> I recently started to read "Data Analysis, A Bayesian Tutorial" by D.S.
> Silva (currently reading chapter 4, unfortunately real work is always
> interfering) and I really like the approach and the straight forward
> manner in which the theory builds up.
Sivia&Skilling's book is very good. A similar and good one is Gregory's
"Bayesian Logical Data Analysis for the Physical Sciences".
> However, I feel that the Bayesian approach, is much more difficult to
> translate to practical methods I can implement, but I may be biased by
> the long term exposition to the "recipe based" Frequentist approach.
My opinion is that the Bayesian approach is so much about modelling
the specific system under analysis that is I don't believe there
is a shortcut or a book of recipes that fits every needs. In other
words every practical application usually has its own peculiarities
that frequently leads to a custom solution.
Having said that, in my experience I frequently relied upon
hierarchical/multilevel modelling that can be approached with
Monte Carlo techniques. In that case the math can be simple
(with caveats) and the hard part is done by the Monte Carlo sampler.
> Can someone suggest me some resources (documentation or code) where some
> practical approaches to Bayesian analysis are taught?
If you come from a frequentist mindset (t-test, ANOVA, etc.) you
might find some quick and interesting thing in this book:
"A Practical Course in Bayesian Graphical Modeling", by Lee and
Wagenmakers. In this book you will find a quick approach to the
hierarchical/multilevel modelling mentioned above. The book
illustrates examples using winBUGS - a sampler that I find puzzling
and that I don't like. Luckily you can do the same with the very good
or by writing you own sampler.
More information about the SciPy-User