Hi R-users I'm having an interesting problem in using MCMCglmm for a meta analysis. I run models as I would normally run them, with diffuse priors on fixed and random effects (fixed: mu=0, V=10e8; random: V=1, nu=0.001; gaussian model), and the posteriors I'm getting out of the models are not like anything I've seen before. The range is very high but the variance is very low, so that fixed effect posteriors are a spike around the mean and random effects posteriors are highly truncated at 0. A handful of coefficient sets are taking extreme values that seem to make no sense at all. I wonder why this is: running for longer does not fix the problem and the chains aren't autocorrelated. Parameter expansion does not help the random coefficients. I'm always seeing a handful of samples that are orders of magnitude larger than the data themselves (which are real weighted mean differences over about -2 to 2) whilst the vast majority are being taken from a much more sensible range. It seems the only solution to getting better posteriors is to make them less diffuse (decrease V for fixed effects, increase nu for random effects). This makes sense, but I'm not comfortable doing it when the posteriors are so sensitive to variances on the priors, and I don't know what it would mean for interpretation of the model. I'm not convinced that a prior can be "too" diffuse, and I'm not sure why these extreme samples are being accepted. But then, perhaps allowing the model to sample from a range that is so much larger than the data just doesn't make sense... although like I say, I would have expected these extreme values to be ditched based on likelihood. If anyone can shed some light, it would be greatly appreciated. For now I'll have to forego some of the random effects and forge ahead with gls... Iain - - - - - - - - - - - - Dr. Iain Stott Environment and Sustainability Institute University of Exeter, Cornwall Campus Tremough, Treliever Road Penryn, Cornwall, TR10 9FE, UK. - - - - - - - - - - - - http://www.exeter.ac.uk/esi/ http://biosciences.exeter.ac.uk/cec/
Can an uninformative prior be too diffuse?
3 messages · Iain Stott, Jarrod Hadfield
Hi Iain, It's a bit hard to diagnose without seeing the data. Would it be possible to post them? Is it possible you have some random terms with very few levels? Cheers, Jarrod Quoting Iain Stott <iainmstott at gmail.com> on Mon, 12 May 2014 11:54:50 +0100:
Hi R-users I'm having an interesting problem in using MCMCglmm for a meta analysis. I run models as I would normally run them, with diffuse priors on fixed and random effects (fixed: mu=0, V=10e8; random: V=1, nu=0.001; gaussian model), and the posteriors I'm getting out of the models are not like anything I've seen before. The range is very high but the variance is very low, so that fixed effect posteriors are a spike around the mean and random effects posteriors are highly truncated at 0. A handful of coefficient sets are taking extreme values that seem to make no sense at all. I wonder why this is: running for longer does not fix the problem and the chains aren't autocorrelated. Parameter expansion does not help the random coefficients. I'm always seeing a handful of samples that are orders of magnitude larger than the data themselves (which are real weighted mean differences over about -2 to 2) whilst the vast majority are being taken from a much more sensible range. It seems the only solution to getting better posteriors is to make them less diffuse (decrease V for fixed effects, increase nu for random effects). This makes sense, but I'm not comfortable doing it when the posteriors are so sensitive to variances on the priors, and I don't know what it would mean for interpretation of the model. I'm not convinced that a prior can be "too" diffuse, and I'm not sure why these extreme samples are being accepted. But then, perhaps allowing the model to sample from a range that is so much larger than the data just doesn't make sense... although like I say, I would have expected these extreme values to be ditched based on likelihood. If anyone can shed some light, it would be greatly appreciated. For now I'll have to forego some of the random effects and forge ahead with gls... Iain - - - - - - - - - - - - Dr. Iain Stott Environment and Sustainability Institute University of Exeter, Cornwall Campus Tremough, Treliever Road Penryn, Cornwall, TR10 9FE, UK. - - - - - - - - - - - - http://www.exeter.ac.uk/esi/ http://biosciences.exeter.ac.uk/cec/
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.
1 day later
Jarrod, you hit the nail on the head. I had a random term with only a couple of levels (taxonomic group) which I was planning to replace with a phylogeny anyway. As soon as I took out that variable from the models, they're running brilliantly. Thanks for your help, and to others that emailed me with similar advice! Iain - - - - - - - - - - - - Dr. Iain Stott Environment and Sustainability Institute University of Exeter, Cornwall Campus Tremough, Treliever Road Penryn, Cornwall, TR10 9EZ, UK. - - - - - - - - - - - - http://www.exeter.ac.uk/esi/ http://biosciences.exeter.ac.uk/cec/
On 12 May 2014 16:09, Jarrod Hadfield <j.hadfield at ed.ac.uk> wrote:
Hi Iain, It's a bit hard to diagnose without seeing the data. Would it be possible to post them? Is it possible you have some random terms with very few levels? Cheers, Jarrod Quoting Iain Stott <iainmstott at gmail.com> on Mon, 12 May 2014 11:54:50 +0100:
Hi R-users I'm having an interesting problem in using MCMCglmm for a meta analysis. I run models as I would normally run them, with diffuse priors on fixed and random effects (fixed: mu=0, V=10e8; random: V=1, nu=0.001; gaussian model), and the posteriors I'm getting out of the models are not like anything I've seen before. The range is very high but the variance is very low, so that fixed effect posteriors are a spike around the mean and random effects posteriors are highly truncated at 0. A handful of coefficient sets are taking extreme values that seem to make no sense at all. I wonder why this is: running for longer does not fix the problem and the chains aren't autocorrelated. Parameter expansion does not help the random coefficients. I'm always seeing a handful of samples that are orders of magnitude larger than the data themselves (which are real weighted mean differences over about -2 to 2) whilst the vast majority are being taken from a much more sensible range. It seems the only solution to getting better posteriors is to make them less diffuse (decrease V for fixed effects, increase nu for random effects). This makes sense, but I'm not comfortable doing it when the posteriors are so sensitive to variances on the priors, and I don't know what it would mean for interpretation of the model. I'm not convinced that a prior can be "too" diffuse, and I'm not sure why these extreme samples are being accepted. But then, perhaps allowing the model to sample from a range that is so much larger than the data just doesn't make sense... although like I say, I would have expected these extreme values to be ditched based on likelihood. If anyone can shed some light, it would be greatly appreciated. For now I'll have to forego some of the random effects and forge ahead with gls... Iain - - - - - - - - - - - - Dr. Iain Stott Environment and Sustainability Institute University of Exeter, Cornwall Campus Tremough, Treliever Road Penryn, Cornwall, TR10 9FE, UK. - - - - - - - - - - - - http://www.exeter.ac.uk/esi/ http://biosciences.exeter.ac.uk/cec/
_______________________________________________ R-sig-mixed-models at r-project.org mailing list https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
-- The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.