 |
 |
A summary of replies from Malin Pinsky <[log in to unmask]> (In general, please post these direct to the list and not me!) ------------- Begin Forwarded Message ------------- Dear Bugs users, The responses I've received to my email this morning have been extremely helpful. The issue was my choice of priors for alpha and beta. Extending them to alpha ~ dnorm(0,1.0E-15) and beta ~ dnorm(0,1.0E-15) gave me answers in line with the answers from traditional stats programs. I've copied the most helpful answers below, and will have another question for the list coming soon. Thanks again. Malin ------------------------- On Friday, October 11, 2002, at 12:24 PM, Kenneth Rice wrote: 2) Your prior for mu isn't as uninformative as it looks. (Try switching it to dnorm(0,1.0E-6) and repeat the analysis). If you're really keen to get the classical estimates out, put flat, dunif, priors on alpha and beta. As long as the the posterior is symmetric, then the medians of alpha and beta should be close to the Excel estimates. Best wishes Ken Rice --------------------------- Hi Malin! The reason for your "strange" results is the prior distribution you assign to alpha. The prior distribution you use has standard deviation of only 1000, which means that your prior distribution is quite informative in this case. Use more flat prior distribution, and you will get the same point estimates as from your least squares analysis. For example, alpha~dnorm(8000,0.0.00000000000001) provides much flatter prior, and will provide prior means very close to the least squares estimates. However, your initial answer is correct, in the sense that it is appropriate compromise between your prior belief and the information in the data. Best regards, Samu Mäntyniemi --------------------------------------------------- It turns out that if you look at the distribution of alpha and beta, you see that they are so wide that the data is obviously not having a strong influence. Thus the priors will tend to dominate, and having beta ~ dnorm(0.0,1.0E-6) Will pull the posterior distribution down even though the standard deviation of this prior is very large. Shifting the priors of both distribtions to a flat prior yields: node mean sd MC error 2.5% median 97.5% start sample alpha 8434.0 1786.0 48.48 4966.0 8407.0 12050.0 1001 10000 beta -122.0 142.6 3.89 -412.5 -121.5 157.8 1001 10000 Which given the standard deviations of these priors is almost exactly what excell gives. I hope this helps and again, sorry for the previous message. Finn Krogstad ------------- End Forwarded Message ------------- From: David Spiegelhalter <[log in to unmask]> Your priors are informative. With a uniform prior on an appropriate range you get the Excel answer. This is a Bayesian program!! David ------------------------------------------------------------------- This list is for discussion of modelling issues and the BUGS software. For help with crashes and error messages, first mail [log in to unmask] To mail the BUGS list, mail to [log in to unmask] Before mailing, please check the archive at www.jiscmail.ac.uk/lists/bugs.html Please do not mail attachments to the list. To leave the BUGS list, send LEAVE BUGS to [log in to unmask] If this fails, mail [log in to unmask], NOT the whole list