MCMC Chains with Large Variance


I am estimating a dynamic IRT model with Rstan. There is no divergence warning but some of the chains have extremely high variance (See the figure attached).

I am wondering what might be the reason behind this issue. The model assumes a Markov process: \theta_{i,t} = N( \theta_{i,t-1}, \sigma_i ). And this surge in variance usually occurs when t is large.

Thank you very much!

unfortunately we can’t help you much with your inquiry without seeing the model code and at least a subset of the data…

EDIT: Should have mentioned that high variance can happen quite easily when the data do not provide enough information about the coefficients. Which may mean the model is OK, but you have insufficient data, but also that the model is a bad match to what you measured (hard to say without knowing more about the model and data)

1 Like