Bootstrapped p-values for LP test

[Cem279]

Hello All,

I need a code calculating bootstrapped p values for Lumsdaine and Papell (1997) test with 4 breakpoints. When I run the code (posted here) I have noticed that Lumsdaine Papell test allows more than 2 breakpoints. In the paper I have been working on one of the series seems to have 4 breakpoints, therefore I am unable to use the significance levels as reported in the original LP paper.

Thanks for any help or suggestions.
Cem.

[TomDoan]

Hello All,

I need a code calculating bootstrapped p values for Lumsdaine and Papell (1997) test with 4 breakpoints. When I run the code (posted here) I have noticed that Lumsdaine Papell test allows more than 2 breakpoints. In the paper I have been working on one of the series seems to have 4 breakpoints, therefore I am unable to use the significance levels as reported in the original LP paper.

Thanks for any help or suggestions.
Cem.

Aren't you getting significant test statistics with one or two breaks? If so, you're done. While we allow for more in the LP procedure (and some of the others), the tests with more than two breaks are unlikely to be useful. You might reject the unit root with one or two then observe that there seem to be more than that, but again, you've already rejected the original null.

[Cem279]

Hello All,

I need a code calculating bootstrapped p values for Lumsdaine and Papell (1997) test with 4 breakpoints. When I run the code (posted here) I have noticed that Lumsdaine Papell test allows more than 2 breakpoints. In the paper I have been working on one of the series seems to have 4 breakpoints, therefore I am unable to use the significance levels as reported in the original LP paper.

Thanks for any help or suggestions.
Cem.

Aren't you getting significant test statistics with one or two breaks? If so, you're done. While we allow for more in the LP procedure (and some of the others), the tests with more than two breaks are unlikely to be useful. You might reject the unit root with one or two then observe that there seem to be more than that, but again, you've already rejected the original null.

Dear Tom Doan.

Many thanks for your prompt reply.The issue is that I am at a loss to determine whether my series are indeed I(0) or I(1) depending on the results from Zivot-Andrews 1 breakpoint test and Lumsdaine-Papell 2 and 4 breakpoints tests.
The results are
Zivot-Andrews:

Variable(s) USA_BETA

t-stat(s) -5.542244
Lag(s) 0.000000
Break 2007M12
DU1 p-value 0.000248
2 breaks. A is chosen
A B C (Significance levels)
1% -5.34 -4.93 -5.57
5% -4.80 -4.42 -5.08
10% -4.58 -4.11 -4.82

Lumsdaine_Papell Test (4 BREAKS):
Lumsdaine-Papell Unit Root Test, Series USA_BETA
Regression Run From 6 to 102
Observations 98
Breaks in Intercept and Trend
Breaks at 25 40 57 77

Variable Coefficient T-Stat
Y{1} -0.8738 -7.9021
Lumsdaine-Papell Unit Root Test, Series USA_BETA
Regression Run From 6 to 102
Observations 98
Breaks in Intercept and Trend
Breaks at 40 77

Variable Coefficient T-Stat
Y{1} -0.7083 -6.2383
Significance levels:
Notes: The critical values are -7.34 (1%),-7.02 (2.5%), -6.82 (5%), and -6.49 (10%). t-statistics are in parentheses.

Now according to ZA test the series is significant therefore stationary, but according to LP with 2 breaks it is not whereas when we extend to 4 breaks it seems to be significant therefore stationary.
Which one to consider ?
Attached with this post please find the graph of the series USA_BETA (lower panel) to give you an idea.
Best Regards.
Cem

[TomDoan]

USA_BETA is dead flat for over half the sample. Clearly none of the DGP's involved in the Z-A or L-P tests is going to be even close to correct. Why do you want to test such a series for unit roots?

[Cem279]

USA_BETA is dead flat for over half the sample. Clearly none of the DGP's involved in the Z-A or L-P tests is going to be even close to correct. Why do you want to test such a series for unit roots?

Dear Tom.

Sorry for the late reply I had a flu which kept me away from my work.

This series is nothing but the “banking-sector beta” which is the standard capital asset pricing model (CAPM) beta, and is defined as follows:

Beta = Cov(rm,rb) / sigmasq.

where rb and rm represent the year-over-year banking or market returns divided by market return variance, computed over a 12-month rolling window.

This series is one of the vital components of financial stress index (FSI) developed by IMF. What I am looking for is the response of emerging markets FSI to USA_beta (American Beta) used as one of the explanatory variables in a Markov Switching context.

My preliminary resulst with a MSW indicate that during tranquil (calm) times (first state) USA-beta is negative and significant ,and during turmoil (state two) it is positive and signficant.
If I can be assured of its stationarity therefore having a fake unit-root due to some level shifts, then it will be quite revealing for me.
Many thanks for a comment on this. If you still suggest that I remove it from the model I will do so.
Regards.

[TomDoan]

USA_BETA is dead flat for over half the sample. Clearly none of the DGP's involved in the Z-A or L-P tests is going to be even close to correct. Why do you want to test such a series for unit roots?

Dear Tom.

Sorry for the late reply I had a flu which kept me away from my work.

This series is nothing but the “banking-sector beta” which is the standard capital asset pricing model (CAPM) beta, and is defined as follows:

Beta = Cov(rm,rb) / sigmasq.

where rb and rm represent the year-over-year banking or market returns divided by market return variance, computed over a 12-month rolling window.

This series is one of the vital components of financial stress index (FSI) developed by IMF. What I am looking for is the response of emerging markets FSI to USA_beta (American Beta) used as one of the explanatory variables in a Markov Switching context.

My preliminary resulst with a MSW indicate that during tranquil (calm) times (first state) USA-beta is negative and significant ,and during turmoil (state two) it is positive and signficant.
If I can be assured of its stationarity therefore having a fake unit-root due to some level shifts, then it will be quite revealing for me.
Many thanks for a comment on this. If you still suggest that I remove it from the model I will do so.
Regards.

You probably need to check with the source of that data series. By your description, it shouldn't have two long stretches with a single (rather implausible) value.