Convergence problems with Rabanal/Rubio-Ramirez model

Hi all,

I’ve encountered some convergence problems with the simple Rabanal/Rubio-Ramirez model when using observed data. Just using the mod-file posted by Stephane’s BVAR-DSGE with the dsge_prior_weight switched off. I’ve used what I think the exact data and detrending procedure as in Rabanal and Rubio-Ramirez (2001) working paper version (Fed Atlanta 2001-22). The underlying model is the simple baseline model without sticky wages and price indexation.

data from Bureau of Labor Statistics, 1982:1 to 2000:4

output: non-farm business output divided by the implicit price deflator, taking logs and removed the mean and detrended with a linear trend.

inflation: quarterly inflation of the implicit deflator of nfb output, demeaned, and detrended (linear), expressed as decimals

interest rate: quartely fedfund rates, demeaned, detrended (linear), expressed as decimals

real wages: hourly compensation of nfb ouput, divided by implicit price deflator, taking logs, demeaned and detrended (linear)

What happened is that even with 2 blocks and each 2.000.000 draws, the Gelman and Brooks plots show that some parameters are actually diverging (inverse sigma, thetabig, rhoy, rhopie, etc.). Also manually computing the Geweke’s CD statistics reveals that most of the parameters didn’t converge. The average acceptance ratios are around 0.30

I’ve even tried with HP filtered series, longer estimation sample and other output/inflation measures like GDP/GDP deflators. The convergence results are all bad.

This is very puzzling since I am presumably using the same data as in Rabanal and Rubio-Ramirez.

Many Thanks, Tomi

Hey yo,

The convergence problem might be due to the priors. If you are using non-informative priors, the model doesn’t work, and all the convergence problems arise.

You should try to reduce the prior standard deviations.

Note that some of the prior distributions used by Rabanal Rubio-Ramírez aren’t exactly the “best choice” that you could try for those parameters. That maybe has something to do with your problem too.