Hi Remco,
Thank you sincerely for the swift reply - ditching the tree logger has been a huge relief, as in the present case I'm not interested in the posterior sample or anything other than the marginal likelihood and standard deviation estimates. The explanation regarding how various parameters interact is also appreciated (I felt the tutorial got most of it across, but missed some key points), particularly the irrelevance of sampling frequency. My chains now get set to ludicrously high values, and I've finally had a 400 particle analysis resolve (previously I said they terminated under 200k steps, this was an error and referred to the 64 particle run, which provided a standard deviation that was too high to distinguish my models).
I have hit a GC overhead limit, but that's (a) not surprising, (b) fixable and (c) irrelevant to my final numbers (this time).
I do have a question regarding the output of a nested sampling run: although the FAQ states that the three estimates of the marginal likelihood provided at the end of the run should not be different enough to worry about (as long as the appropriate SD is reported), I'm still puzzled by the difference in how each is generated - one has the SD marked "subsample", another "bootstrap", and the third is presented simply with brackets around the standard deviation value.
Assuming it can be explained in broad terms, how are these three SDs produced, as distinct from yet another SD value provided below? Also, which of the marginal likelihood estimates is supposed to match that fourth SD (printed under information content)? Or is that SD supposed to accompany the final value printed to screen before the conclusion of the run?
e.g. (from an earlier run)
Marginal likelihood: -23154.59891321004 (bootstrap SD=5.099546305106115)
Marginal likelihood: -23154.426732159405 (subsample SD=4.916936211187053)
Marginal likelihood: -23154.964067192115(4.375577431226804)
Information: 1440.6436677045785
SD: 4.744476505146173
On the one hand, these values are all clearly within spitting distance of one another, so it will make no difference to my model selection (the desired outcome) - on the other hand, I can see how I may need to describe the values in a paper, or explain to reviewer why I selected a particular value. It might not matter from an immediate practical standpoint, but it might end up mattering from a publication standpoint (which is another kind of practical standpoint).
Thanks again for the help,
Cheers,
-Kate