I am trying to run jpfeifer’s Garcia-Cicco et al (2010) code with different data set, but I keep on getting very narrow posteriors (especially for the measurement error se’s). I assumed this was due to the wrong initial values, but if I make a slight change in the initial value, then I end up with even worse results like bimodal posteriors.
I tried putting the peak or mode of these posteriors as the initial value for each parameter, but it wouldn’t get much better. I also tried different mode check options like 9, but the “error using chol” says that “the matrix must be positive definite.”
What should I do to improve the posteriors to be more normal shaped? Does it really have something to do with the specification of initial values? Or are the posteriors good enough already? Any advice would be greatly appreciated!
Please check the convergence diagnostics. I guess you need a longer chain. Regarding the prior posterior plots: with such a diffuse prior they can seem misleading. Please have a look at the HPDIs reported after estimation to see whether they really are too narrow.
Thank you very much for your reply. I don’t want to bother you with too basic questions but I couldn’t figure out these anywhere else.
Could you check if I’m doing the right thing: I used your initial values from the original .mod file as a starting point, and then for the second and third run, I put the peak of the previous posteriors as new initial values. For the third run I increased mh_replic from 350000 to 700000. The attached files are the result of the third run. Is it normal/necessary that I run the code several times like I did? Should I try the fourth run?
Also, would you say the measurement error standard error’s HPDI length around 0.0005 is long enough?