On Feb 14, 2014, at 10:23 AM, Christopher Green wrote:

On 2014-02-13, at 10:38 AM, Paul C Bernhardt wrote:


My only problem with the Bayesian approach, described elegantly in the article, 
is that the posterior probabilities are so dependent on the prior probabilities.


I hear this all this time, but I disagree. Even wildly divergent priors 
converge fairly rapidly in the face of the same data. In any case, priors are 
not necessarily reflective of mere "bias." They are often reflective of true 
expertise that has been developed informally in a field. To use a classic 
example, if you want to know what the probability is that two countries will go 
to war over, say, the next decade, do you think it is better to start with the 
base probability that ANY two random countries will go to war, or would you 
rather start with the estimate of people who are already expert in the history, 
cultures, and economies of the two countries in question? Bayesian prior, when 
used properly, actually SHORTEN the time to decision, not distort it.

I agree, and that's fine and good, technically.

But, there is a major practical problem we all face, that replications are not 
generally publishable, and even if publishable, is not the stuff of which 
tenure is made at Tier 1 Research Universities. So, you have one study on which 
to show your phenomenon is real. It has to be non-obvious enough to get you 
down the tenure track, but known well enough that prior probabilities are 
decently high… oh, and be something that you are interested in, for which data 
can be gathered in a short enough time period, etc.

Good luck with finding that unicorn-esque project, much less executing it.

The culture is not just a wide-spread attitude. The culture is in synergistic 
relationship with systems that discourage use of anything but traditional 
statistical thinking.

When such a tightly integrated system exists, changing it is nigh impossible. 
What will cause it to change is one of two things: 1) A catastrophic 
failure/scandal such that the entire enterprise needs to be overhauled (e.g., 
the appalling use of research participants that leads to the Belmont Report). 
That seems unlikely because it will be too difficult to show lives destroyed by 
misapplication of statistical procedures/approaches. (I wonder if that was 
Bem's goal with the precognition paper, to provide results so bizarre through 
routine statistical approach that it would embarrass and 'shake up the 
business'.) 2) Psychology being supplanted by other disciplines that use more 
sophisticated methods and begin to make greater advances and have greater 
practical effect. At a certain point people may realize we are being made 
irrelevant and after years of anxious hand-wringing some kinds of fixes will be 
implemented, hopefully not too late.

Or, we can fix it now.

If I were editor of a journal I would simply require parallel results and 
discussion sections. One traditional in approach, the other Bayesian in 
approach. This would allow folks who are embedded in the past to continue their 
work the way they always have, but allow those who want to move to new 
approaches to do so. Eventually, if the Bayesian approach yields more useful 
interpretations of the data, it will become apparent and we should see changes 
occur in what is taught to students. As soon as there is a predominance of 
researchers who are Bayesian in approach the shift will be complete and we will 
see journals dropping the requirements for traditional approaches. We will all 
be dead before this happens, of course.

Paul

---
You are currently subscribed to tips as: [email protected].
To unsubscribe click here: 
http://fsulist.frostburg.edu/u?id=13090.68da6e6e5325aa33287ff385b70df5d5&n=T&l=tips&o=34207
or send a blank email to 
leave-34207-13090.68da6e6e5325aa33287ff385b70df...@fsulist.frostburg.edu

Reply via email to