Dear everyone, dear Christopher,
A few technical questions about the inner workings of OxCal:
1. When I am adjusting a date by offsetting it with a probability distribution such as Exp(-1, 0, 300), am I correct in thinking that the MCMC sampler samples from the whole range [0, 300]?
The reason I am wondering about it is because when I look up the "Raw Data", the numbers printed with their probability densities do not cover the whole range (see picture below). But the effect of adding such an offset (to model possible old wood) does seem to be such as if the sampling actually explores the possibility of adding e.g. +200.
1a. What exactly is the Posterior in that Raw data display? (In some of my examples, it appears that the actual offset applied is much greater than the maximum numbers printed out in this table; hence the question.)
2. Is there a way to force the MCMC sampler to run for a given _minimum_ number of cycles? The Maximum argument of MCMC_Sample() really defines the maximum, which does not have to be reached.
The reason I’m asking is because in some cases, running a specific analysis for longer might be beneficial. Sometimes, OxCal sampling results vary between independent runs - in my experience so far, the variation is small, on the order of 5-10 years for the 95% interval boundaries. But still, sampling the posterior for longer may lead to a better estimate (I would not want to do this for every run, but when I am happy with a model and want to prepare its "final and clean" run, this could be nice to do.) Obviously the "problem" is not very serious, it’s more about having "nice", marginally cleaner, results.
Furthermore, in one real-life analysis I did in the past, there was an anomaly in the outputs which seemed to stem from inefficient traversing of the posterior. (I don’t remember the details now, but I think the problem was that the model effectively defined a ridge of two correlated quantities and it was hard for OxCal to explore that ridge in its entirety. I did not check back then whether I could recode the model with a different parametrization to sort it out.) However, in such cases simply running MCMC for longer might not help, so I’m mainly asking for the first type of case.
3. A similar but more technically complex matter: can one force OxCal’s MCMC to sample some parameters more frequently than others?
In the analysis I am working on at the moment, some distributions are plotted in a "toothed" manner, which suggests insufficient exploration of the relevant distribution (picture attached below - the "teeth" are in the prior distribution in this case.)
Thank you in advance for the answers!
All the best,
Igor
Raw Data view for Exp(-1, 0, 300):
"Teeth":
