This question is not about a specific model, therefore I’m not posting the model here. I’m confused about some of the details of the reports and results from Stan fits in general.
This is a report header from a recent Stan fit, which was set up to run 4 chains, iterations=50000, warmup=5000:
‘’’ Stan report
4 chains: each with iter=(50000,50000,50000,50000); warmup=(0,0,0,0); thin=(1,1,1,1); 200000 iterations saved.
Warmup took (1192, 1168, 1138, 1155) seconds, 1.3 hours total
Sampling took (10861, 8264, 10847, 6912) seconds, 10 hours total
I have a few questions:
- The report says 200k iterations were saved, but I see only 50k iterations in the results. Are the results from different chains averaged, or something else going on?
- Why does the report say warmup=0? I passed on 5k warmups when I started the fit and Stan didn’t complain (i.e. proceeded to run and completed the fit). And Stan clearly spent about 10% of its time on the warmup, which suggests it carried out 5k warmups for 50k iterations as I asked.
- Are warmup iterations saved anywhere? Is it possible to see how the spin-up (this may not be the right phrasing, sorry) distributions look? This is important because I do not see a clear path to determining the proper number of warmups for a given model unless one can compare the warmup results to the “real” iterations.