F test vs. mcmcpvalue
On 08/07/2008, at 3:16 AM, Ben Bolker wrote:
Hank Stevens wrote: d> HI Ben and Spencer, | Thank you very much for your help. | | 1. The QQ plots look normal, but highlight the lack of balance (from one | to dozens of reps per treatment combo). | 2. The MCMC sample traces look (in my limited experience) without | peculiarities, and the densityplots are all quite symmetrical and | normal-ish. | 3. Simulations (lmer::simulate) of the null hypothesis indicate that | F-stats as large (or larger) than my observed F-stats are VERY unlikely, | under the null hypothesis. | | As I learn anything else useful, I will be happy to share. | Cheers, | Hank | ~ #3 pretty much seals it for me -- since that is really what the F test is trying to test. ~ It's a little hard to reconcile #2 and #3, though ... I would think you could move on at this point, but just for laughs -- are you using mcmcpvalue on a single contrast, or multiple parameters? If the former, does it seem to agree with the results of HPDinterval() or quantile()? If the latter, is there something about the _combinations_ of parameters that is wonky?
If an MCMC isn't traversing the parameter space properly the traces will probably still look OK until it shifts into a new region which may take a while. Also it was mentioned that there were 500 observations. For clustered data it is the number of clusters that is more important. Ken