Hi, I have a parameter (lkj_corelation of 2x2) that has high R-hat = 1.05. I ran 1500 warms + 1500 iterations for 4 chains.
I plotted out the traceplot, which looks like the following. Is this suggesting auto-correlation in the draws? would this be a big concern? The exact value of R-hat for L_sigma[2,1] and L_sigma[2,2] is 1.04, and 1.05 respectively.
Would increasing sample size help?
The biggest concern is the bimodality of the distirbution which suggests that the correlation is degenerate. It may be that the relationship of your data or parameters are not being well-captured by this dependency structure, or that your data has a nonlinear relationship that leads to apparent correlation coefficients at both positive and negative values. This may either prescribe using a different correlation structure or sorting your data such that conditions are properly matched, depending on if this is a true nonlinear correlation or an artifact of including data you shouldn’t in the comparison.
That would be reasonable to test and would not take much time. If the effective sample size (ESS (you seem to be using an old interface with N_eff)) doubles when the total sample size doubles then it’s likely that all is good and the chains just have high autocorrelation (probably due to highly varying curvature making the fixed step size NUTS to be slow)
