Bai-Perron JAE 2003 Replication Files

[TomDoan]

The attached zip includes the two empirical examples from Bai & Perron(2003), "Computation and analysis of multiple structural change models," Journal of Applied Econometrics, vol. 18, no. 1, pages 1-22. The U.S. real rates example is a regression on constant alone, while the UKPhillips example does an AR(1) model and a Phillips curve with two fixed regressors.

BaiPerronJAE2003.zip

(2.9 KiB) Downloaded 487 times

The new version of the BaiPerron procedure needed by this is at:

viewtopic.php?f=7&t=500

[jkapsoli]

When I tried to replicate the paper I found the following error message:

## SX9. REPORT is not a Basic Data Type
>>>>type report <<<<
## I2. Expected Instruction Here

I'm using version 7.00 is that the problem?

[TomDoan]

LOCAL REPORT was added with RATS v7.10. There's really no way to get rid of that in the reworked baiperron.src because it works on two separate reports (for the breaks and for the regressors). You can patch a version 7 to 7.10 using http://www.estima.com/resources_patches.shtml. Even better, you should be able to get a 7.3 from your IT department.

[roger02]

Dear Tom
is the bai-perron sensitive to the number of structural breaks? i have weekly data of 702, what would be the maximum number of breaks? second, how can i test for the confidence intervals for each break?
thanx

[TomDoan]

It slows down markedly when the number of breaks goes up. One question to ask is whether it makes any sense to have a "model" with say, five structural breaks.

[TomDoan]

A revised version of BAIPERRON.SRC has been posted at:

http://www.estima.com/forum/viewtopic.php?f=7&t=500

This includes confidence bands on the break points.

After working through the various options for computing those, I would caution anyone about putting too much credence in them. They can be very sensitive to minor differences in choice of lag windows for computing long run variances. Also note that they do not take into account the actual difference in sums of squares in moving the break points around by a few entries; instead, they are based on a prediction of the local change based upon the difference in coefficients between the two partitions on either side.

[nacrointfin]

Hi Tom:

Can the test of Kejriwal, M. and P. Perron (2009). "Testing for Multiple Structural Changes in Cointegrated Regression Models." Journal of Business and Economic Statistics be implemented by the code of Bai-Perron JAE 2003?

Regards,

Terence

[TomDoan]

Hi Tom:

Can the test of Kejriwal, M. and P. Perron (2009). "Testing for Multiple Structural Changes in Cointegrated Regression Models." Journal of Business and Economic Statistics be implemented by the code of Bai-Perron JAE 2003?

Regards,

Terence

That's correct. The number crunching is the same as is done by the BaiPerron procedure. The 2009 paper derives the asymptotic distribution of the break test statistics under assumptions allowing for cointegration.

[iloverats]

The attached zip includes the two empirical examples from Bai & Perron(2003), "Computation and analysis of multiple structural change models," Journal of Applied Econometrics, vol. 18, no. 1, pages 1-22. The U.S. real rates example is a regression on constant alone, while the UKPhillips example does an AR(1) model and a Phillips curve with two fixed regressors.

BaiPerronJAE2003.zip

The new version of the BaiPerron procedure needed by this is at:

viewtopic.php?f=7&t=500

dear
why are the results form the Rats so different form the paper?


results form the Rats
Breaks RSS BIC LWZ F(m) F(m|m-1)
0 1214.9219 2.51 2.55
1 644.9955 1.92 2.00 89.24 89.24
2 455.9502 1.62* 1.74* 83.23 41.46
3 445.1819 1.64 1.79 57.06 2.39
4 444.8797 1.69 1.88 42.41 0.07
5 449.6395 1.74 1.97 33.02 -1.03

paper results:
supf(1) =57.91
supf(2)= 43.01
supf(3)= 33.22
supf(4)=24.77
supf(5)=18.33
....
which is right?

[TomDoan]

Both. The results in the paper use F statistics computed with HAC covariance matrices using a very specific procedure for estimating the long-run variance. Computing that requires a great deal of number-crunching that the dynamic programming algorithm is trying to avoid, but which is feasible here with a relatively small data set. RATS is using the simpler standard F. They both give the same results asymptotically.

[iloverats]

Both. The results in the paper use F statistics computed with HAC covariance matrices using a very specific procedure for estimating the long-run variance. Computing that requires a great deal of number-crunching that the dynamic programming algorithm is trying to avoid, but which is feasible here with a relatively small data set. RATS is using the simpler standard F. They both give the same results asymptotically.

if i use Rats , are the critical values in their papers(1998)" Estimating and Testing Linear Models with Multiple structural Changes" ok?
thank you very much

[TomDoan]

Both. The results in the paper use F statistics computed with HAC covariance matrices using a very specific procedure for estimating the long-run variance. Computing that requires a great deal of number-crunching that the dynamic programming algorithm is trying to avoid, but which is feasible here with a relatively small data set. RATS is using the simpler standard F. They both give the same results asymptotically.

if i use Rats , are the critical values in their papers(1998)" Estimating and Testing Linear Models with Multiple structural Changes" ok?
thank you very much

Yes. Both the HAC and the standard F are covered by the same theorem that lays out the asympotics.

[Barry Quinn]

Both. The results in the paper use F statistics computed with HAC covariance matrices using a very specific procedure for estimating the long-run variance. Computing that requires a great deal of number-crunching that the dynamic programming algorithm is trying to avoid, but which is feasible here with a relatively small data set. RATS is using the simpler standard F. They both give the same results asymptotically.

Apologies for digging up an old post but my query is directly related to this answer.
In @baiperron is it possible to obtain HAC adjusted F statistics by changing the robust part of this code to 1:
line 274-285

Code: Select all

procedure BPBreakRanges startr endr eqnshift limits
type integer       startr endr
type equation       eqnshift
type rect[int]    *limits
*
option vect[int]  breaks
option integer    maxbreaks
option integer    nfix
option integer    nshift
option switch     robust         0
option switch     qhet            0
option switch      omegahet         0

When i trim my data the sample is relatively small at 50 observations and i would like to compute the HAC covariance matrices as per original paper
This is probably too simplistic a solution to this complex calc!!

thanks

[irfansystem]

I do not understand the F(m) and F(m|m-1)....what do they exactly represent??

Breaks RSS BIC LWZ F(m) F(m|m-1)
0 1214.9219 2.51 2.55
1 644.9955 1.92 2.00 89.24 89.24
2 455.9502 1.62* 1.74* 83.23 41.46
3 445.1819 1.64 1.79 57.06 2.39
4 444.8797 1.69 1.88 42.41 0.07
5 449.6395 1.74 1.97 33.02 -1.03

is there a way to get values for SupFt(1)........UDmax and WDmax? as given in the Bai Perron (2003) page 59.

PS: I do realize that F(m) and F(m/m-1) has got to do something with SupFt but it is not clear to me what is what actually!!

[Ana_Rita]

Check this paper by them "Estimating and testing linear models with multiple structural changes" source Econometrica, vol.66, no. 1 (Jan, 1998), pp.47-78.

They show tables of their estimated critical values for the tests. I did try and mach them with the results they present in their 2003 paper but with no succes though...maybe you'll be more lucky then me!

Good luck!