People keep telling me that Sas isn’t as bad as everybody says, but then I see (from Christian Robert) this listing from the Sas website of “disadvantages in using Bayesian analysis”:
There is no correct way to choose a prior. Bayesian inferences require skills to translate prior beliefs into a mathematically formulated prior. If you do not proceed with caution, you can generate misleading results. . . . From a practical point of view, it might sometimes be difficult to convince subject matter experts who do not agree with the validity of the chosen prior.
That is so tacky! As if least squares, logistic regressions, Cox models, and all those other likelihoods mentioned in the Sas documentation are so automatically convincing to subject matter experts.
P.S. For some more serious objections to Bayesian statistics, see here and here.
P.P.S. In case you’re wondering why I’m commenting on month-old blog entries . . . I have a monthlong backlog of entries, and I’m spooling them out day by day. I actually wrote this one on 7 May.
A danger with old blog entries is that they appear with old dates in your RSS feed (this post comes up as from May 7 on your feed). It's possible that some feed readers may expire old posts.
Edward: I asked a colleague, who writes:
Andrew, I have the same issue with old posts showing up far back in the RSS feed. I use Thunderbird as my RSS reader so I can see my blogs in the same place as I'm seeing my emails… I don't believe that there is a choice of what date to use in the preferences anywhere.
Specifically, the RSS feed has the following:
<published>2010-06-13T13:26:30Z</published>
<updated>2010-05-07T09:34:16Z</updated>
So it was "updated" *before* it was published. I would consider this to be a bug in the blogging software, but I think maybe my own software (wordpress) does the same thing…
I suggest perhaps you at least make some minor modification and save the draft right before clicking publish.
Does it happen all the time? I ask because nowadays most of my blog entries are published about a month after I write them. The blog has a Scheduled feature that allows me to do it.
For example, I wrote this one awhile ago, and also this one.
Hi Andrew
I like SAS, so let me say a few things (as an aside, it's SAS not Sas).
First, I heard a talk by Oliver Schabenberger of SAS regarding the Bayesian procedures, and I thought it was a fairly balanced presentation of frequentist vs. Bayesian views.
Second, I think that it often IS hard to convince substantive experts about Bayesian analysis, and, on the other hand, in my experience, substantive experts are far too willing to accept what the frequentist analysis says, and reluctant to give the statistician the substantive knowledge that is necessary to perform the best analysis, whether frequentist or Bayesian.
Peter I would second your second – but it would have been better if the documentation had in balanced fashion raised the same concerns about the data model (likelihood) and there being no correct way to choose that either
For an example of the damage an uncritical use/acceptance of the "Cox model" (proportional hazards) resulted in see here
http://www.stat.columbia.edu/~cook/movabletype/ar…
Keith
Keith:
Exactly. I really don't like the cringing approach in which any classical procedure gets a free pass and anything Bayes gets all this extra scrutiny. If you're encouraging people to think of likelihoods as God-given and infallible, then I think you're part of the problem, not part of the solution.
Hi Keith and Andrew
I think our differences are that I am trying to describe (and I think the quoted text in Andrew's original post was, as well), while both of you may be trying to prescribe. Certainly classical approaches are overused – I've written extensively on MEDSTAT and elsewhere about the damage cause by blindly following significance testing – but, at least in my experience, that is how it is. People DO accept what the traditional analysis says, even when they shouldn't.
I fully agree that there is (certainly in the social sciences) very rarely, if ever, one RIGHT model, one TRUTH. But my clients tend to insist on that.
Peter
Peter:
They write, "Bayesian inferences require skills to translate prior beliefs into a mathematically formulated prior. If you do not proceed with caution, you can generate misleading results."
The implication, I believe, is that with non-Bayesian methods, that there is not such a risk of "generating misleading results." I disagree with that implication and I don't think it's accurate as a description (or as a prescription).
Andrew
Good point. I do think that's the implication, and I do agree it's wrong. There's that danger with any type of analysis.
Peter
I like what David Mackay said in his book on Information Theory (p. 345): "This answer depends on several subjective assumptions; in particular, the probability assigned to the free parameters n, c, d, e of the theories. Bayesians make no apologies for this: there is no such thing as inference or prediction without assumptions."
James:
David MacKay is great, and I agree with the "no apologies" attitude.
But I think in the quote you give, MacKay displays the occasional over-enthusiasm that outsiders occasionally feel about particular statistical methods.
Yes, there is no inference without assumptions, but, no, prior distributions are not always required for inference. There is a wide range of statistical problems for which you can get inference, and even calibrated prediction, using a data model and no prior distribution.
At the same time, I feel that MacKay (at least in your quoted excerpt; I don't know the context) is giving up too much by describing the prior distribution as "subjective" while not recognizing the corresponding subjectivity of the data distribution.