Thank you for taking the time to answer in such detail, @OriolAbril , this is all very useful for me to get going again with my models!
This is a core point I had to get my head around.
The original reason for me to draw less samples was a performance issue, which was due to a shape limitation when using LKJ priors. This has been solved in the meantime, prediction works in reasonable time, and my comparison models are happily sampling.
Technical note: the following line is a bit enigmatic to me:
thinned_idata = idata.sel(draw=slice(None, None, 5))
I am not so much into xarray syntax and had little use for slice except for xarray indexing; i learned this once in a scipy conference workshop but keep forgetting due to lack of use.
In the case of model inference data thinning, I found the second argument of slice to be the one to go. But that is just trying around. Might be worth porting some basic xarray documentation to the pymc or arvic api docs?
Thanks again!
Cheers!