Raftery and Lewis (1992) convergence only with one mcmc block

Estimation ML/Bayesian estimation
1

Hi,

I found out that the Raftery and Lewis (1992) convergence test (called with “raftery_lewis_diagnostics” in the estimation command) is not implemented when “mh_nblocks=2”. Is there a reason for that? I know that when there are two chains, the Brooks and Gelman tests are possible, but why does it mean that the Raftery and Lewis test will not be implemented, since it was requested in the estimation command? It is all in the function ‘McMCDiagnostics.m’.

Yaakov

2

I guess we did this because the interpretation is not straightforward. With multiple chains, you could test for convergence of each chain using the Raftery and Lewis (1992) diagnostics. But you would not be using the information across chains. For example, the two chains could each “converge” to a different mode and you would not detect the problem.

2 Likes