Re: [Bugs-r-devel] R2WinBUGS bug fix

[Uwe L. <li...@st...>]

Gorjanc Gregor wrote:
> A better idea would be top write a sample() method for
> bugs class. This would be the most natural way in my
> opinion.


Sure, that would be much cleaner code, but it might break code of people 
using the package. Even if the original design decision is R-unlike (and 
it was my fault not to change it - and insist on such a change - while 
moving Andrew's code into an R package), we should stay with it, I 
think. Note that it is also documented for quite a long time now. ?bugs 
tells us:

sims.matrix     matrix of simulation output, with n.chains*n.keep rows 
and one column for each element of each saved parameter (for 
convenience, the n.keep*n.chains simulations in sims.matrix and 
sims.list (but NOT sims.array) have been randomly permuted)

Best,
Uwe




> sample(bugs)
> ...
> ________________________________________
> From: bug...@li... [bug...@li...] On Behalf Of Andrew Gelman [ge...@st...]
> Sent: Thursday, October 04, 2007 12:55 AM
> To: Anon.
> Cc: bug...@li...
> Subject: Re: [Bugs-r-devel] R2WinBUGS bug fix
> 
> I'm confused because bugs has sims.list, sims.matrix, and sims.array,
> but I don't see sims.  In the ?bugs help, it says that "for convenience,
> the 'n.keep * n.chains' simulations in sims.matrix and sims.list (but
> NOT sims.array) have been randomly permuted"
> 
> The reason why we scramble the simulations is so we can more easily
> access random draws from the posterior distribution.  We can then just
> use the first 10 simulations or the first 100 or whatever to do
> posterior simulations (see chapter 6 of BDA for various examples).
> 
> Checking for convergence--which is the same thing as assessing
> mixing--isn't a problem since we check convergence on the unscrambled
> series that are in sims.array.  You can always use sims.array if you
> want to work with these.
> 
> Andrew
> 
> Anon. wrote:
>> Uwe Ligges wrote:
>>> I think this was proposed by Andrew Gelman in the original code and I
>>> already had the discussion but forgot the outcome (people would like
>>> to have it randomized?, Andrew?).
>>>
>>> Anyway, a fix would be to add drop = FALSE as in
>>>
>>>    sims <- sims [sample(n.sims), , drop = FALSE] # scramble (for
>>> convenience in analysis)
>>>
>>>
>> I'm curious - how can you check for convergence if the order is
>> randomised?  I'm thinking of cases where both chains are moving in the
>> same direction towards the part of the posterior with high probability
>> mass.  It also makes mixing more difficult to assess, although this is
>> less critical.
>>
>> I would certainly prefer not to have this randomised - there are
>> disadvantage, and if you want the results randomise, it's easy to do
>> (hey, it's one line of code!).  OTOH, if it's already randomised, it's
>> difficult to impossible to unscramble it.
>>
>> Bob
>>
> 
> --
> Andrew Gelman
> Professor, Department of Statistics
> Professor, Department of Political Science
> Director, Applied Statistics Center
> Columbia University, New York
> ge...@st...
> www.stat.columbia.edu/~gelman
> 
> Usual schedule:
> 
> Mon, Tues, Fri in Statistics Dept ofc:
>   Social Work Bldg (Amsterdam Ave at 122 St), Room 1016
>   phone 212-851-2142, fax 212-851-2164
> Wed, Thurs in Political Science Dept ofc:
>   International Affairs Bldg (Amsterdam Ave at 118 St), Room 731
>   phone 212-854-7075, fax 212-222-0598
> Wed, Thurs from 2:30pm onward in Playroom:
>   International Affairs Bldg, Room 707
> 
> I'm out of town 11-12 Oct, 26 Oct, and 15-23 Nov.
> 
> 
> -------------------------------------------------------------------------
> This SF.net email is sponsored by: Splunk Inc.
> Still grepping through log files to find problems?  Stop.
> Now Search log events and configuration files using AJAX and a browser.
> Download your FREE copy of Splunk now >> http://get.splunk.com/
> _______________________________________________
> Bugs-r-devel mailing list
> Bug...@li...
> https://lists.sourceforge.net/lists/listinfo/bugs-r-devel
> 
> -------------------------------------------------------------------------
> This SF.net email is sponsored by: Splunk Inc.
> Still grepping through log files to find problems?  Stop.
> Now Search log events and configuration files using AJAX and a browser.
> Download your FREE copy of Splunk now >> http://get.splunk.com/
> _______________________________________________
> Bugs-r-devel mailing list
> Bug...@li...
> https://lists.sourceforge.net/lists/listinfo/bugs-r-devel