Westerlund error-correction-based panel cointegration tests

[Ermiyas Gebrie]

Dear all,
I am trying to do panel cointegration test using xtwest command or using Westerlund(2007) error correction based panel cointegration test. Because of the cross-sectional dependence of our panel data I used the bootstrap option. But, when I use this command "xtwest debtgdp export , constant trend lags(1) leads(1) lrwindow(3)bootstrap(200)" in the same stata version in different time gives me different Robust P-value. Even, when I use Stata version 12.1 and Stata version 14.2 gives me totally different results.
I have written the result of Stata version 12.1 that I run in a different time and the same command. if you could provide me with some better suggestions on how to approach such a problem and what is the problem with it.
The first estimation

xtwest export import , constant trend lags(1 2) leads(0 1) lrwindow(2)bootstrap(10)

Bootstrapping critical values under H0..........
Calculating Westerlund ECM panel cointegration tests..........

Results for H0: no cointegration
With 41 series and 1 covariate
Average AIC selected lag length: 2
Average AIC selected lead length: 1


Statistic    Value     Z-value    P-value   Robust P-value 

Gt       -2.128      1.880     0.970         0.100      
Ga       -3.983      7.560     1.000         1.000      
Pt      -18.004     -5.120     0.000         0.000      
Pa       -7.207      1.802     0.964         0.100

The second estimation

xtwest export import , constant trend lags(1 2) leads(0 1) lrwindow(2)bootstrap(10)

Bootstrapping critical values under H0..........
Calculating Westerlund ECM panel cointegration tests..........

Results for H0: no cointegration
With 41 series and 1 covariate
Average AIC selected lag length: 2
Average AIC selected lead length: 1


Statistic    Value     Z-value    P-value   Robust P-value 

Gt       -2.128      1.880     0.970         0.200      
Ga       -3.983      7.560     1.000         1.000      
Pt      -18.004     -5.120     0.000         0.000      
Pa       -7.207      1.802     0.964         0.200

Thanks,
Ermiyas

[William Lisowski]

Bootstrapping is a random process. Because you (apparently) did not use set seed to initialize the random number generator to the same seed value before each estimation, you likely had different streams of random numbers for the two estimations and thus obtained different results. Something like the following should, I believe, produce the same output each time you run the commands.

set seed 7334
xtwest export import , constant trend lags(1 2) leads(0 1) lrwindow(2) bootstrap(10)

Do read the output of help set seed for further instruction in setting random number generation seeds to produce reproducible results.

With that said, given that your robust p-values are multiples of the inverse of the number of bootstrap samples (10), I think you are going to need a much larger number of bootstrap samples to get more precise p-values.

[Ermiyas Gebrie]

Thank you very much, William Lisowski, for your help. It works well and I got a similar result after followed your help.