Lanne-Lutkepohl JMCB 2008 (Identification by volatility)

[TomDoan]

Revised 9 August 2012

This is a replication of Lanne and Lutkepohl(2008), "Identifying Monetary Policy Shocks via Changes in Volatility", JMCB, vol 40, no 6, 1131-1149 which identifies an SVAR using variance regimes. The authors use three variance regimes, only two of which use the structural model for the covariance matrix. The third (actually the first in order) is left to be freely estimated. It should be relatively easy to adapt to different numbers of variance regimes.

This has the actual data used. However, there was an error in the coding used in the paper. This is one of those errors which is a good example of why you shouldn't use Gauss (or Matlab) for time series; the data matrix shifted as a result of lags, but the break points didn't adapt. With RATS, the data stay put and the lags are extracted from them.

The second program replicates the "Chow test" results for breaks in the (non-structural) VAR. The published results were done with incorrect break points (off by 13 relative to the description).

llhetero.rpf

Estimation program

(8.56 KiB) Downloaded 1667 times

covbreaktest.rpf

Covariance Break Test Program

(4.37 KiB) Downloaded 1464 times

bmdata6596.txt

Data file

(27.07 KiB) Downloaded 1495 times

[sleu123]

I am using the two estimation programs provided in the forum to replicate Lanne and Lütkepohl’s 2008 paper

I was wondering whether the D2(i) and D3(i) (i=1, 2, 3) parameters in the ‘llhetero’ program are the omegas in Table 2 in the paper. Because the RATS estimated parameters values for D2(i) and D3(i) do not match those listed in the table contained in the erratum released by the original authors.

I am also having difficulty with the impulse response functions. I have assumed that the final “policy” column relates to the IRFs presented in the paper, but these do not appear to coincide with those reported in the paper. In order to compare these IRFs directly to the published versions, I have uncommented the relevant sections of code and run VARIRF for each of the 5 estimated models.

Any help would be greatly appreciated.

[TomDoan]

If you download the new copy, I've added comments which describe the differences in the parameterizations. It's not 100% clear from the paper what form of VAR they use for doing the impulse responses, but it appears that it's the OLS estimates of the full six variable VAR, with only the shock taken from a particular model. I replaced the IRF calculation from the earlier version with one that appears to come up with the "BR" model responses.