Good day,
I have tried to replicate Table 8 of the paper "Dynamic panels with threshold effect and endogeneity" by Seo and Shin (2016) Journal of Econometrics Volume 195, Issue 2, December 2016, Pages 169-186.
I obtained the results below but could not match them with those reported in the paper. I am particularly interested in the estimates reported in Table 8 of their paper.
Any help, please? I do have the code and the data if that helps.
Thanks,
ozey
x = c~Tq~d; @ set regressors (c=cash flow, Tq=Tobin, d=debt)@
/------ output based on analytic variance formula -----/
proportion of upper regime: 0.1958
estimates and s.e. (threshold~lower regime (lag y,x)~delta (1,lag y,x)
0.3590 0.3875 0.0519 -0.0038 -0.1350 0.4409 -0.0594 -0.0039 0.6580
0.0396 0.0919 0.0387 0.0023 0.0466 0.2601 0.0521 0.0019 0.1322
upper regime estimates and s.e. (lag y,x)
0.8284 -0.0075 -0.0077 0.5229
0.2291 0.0457 0.0033 0.1311
Long-run parameter estimates and s.e.
0.0519 -0.0038 -0.1350 -0.0075 -0.0077 0.5229
0.0696 0.0033 0.1239 1.3404 0.2641 2.3207
Overidentification J-statistic (p-value)
45.0039 0.0162
/------ in case of averaging -----/
proportion of upper regime: 0.1940
estimates and s.e. (threshold~lower regime (lag y,x)~delta (1,lag y,x)
0.3601 0.3941 0.0592 -0.0034 -0.1215 0.4091 -0.0770 -0.0035 0.6133
0.0426 0.0869 0.0416 0.0021 0.0441 0.2417 0.0491 0.0018 0.1224
upper regime estimates and s.e. (lag y,x)
0.8032 -0.0179 -0.0069 0.4918
0.2127 0.0440 0.0032 0.1234
Long-run parameter estimates and s.e.
0.0592 -0.0034 -0.1215 -0.0179 -0.0069 0.4918
0.0721 0.0031 0.1118 1.0912 0.2213 1.5592
Overidentification J-statistic (p-value)
64.0851 0.0001
elapsed time : 15 minutes 56.92 seconds
2 Answers
0
There will be some variation in the output of the code because it uses random numbers in the computations. I ran the code in GAUSS 18 on a Mac and GAUSS 10 on a Linux server and got slightly different results each time.
Then I made two changes which gave me reproducible results.
First change
I set the random number seed, using the rndseed keyword towards the top of the code, like this
new; format /rd /m1 8,4; output file= invest.out on;
tstart = date;
load invest; // Load data and define variables
// Set seed for repeatable random numbers
rndseed 23423;Second change
When running the code with GAUSS 18, I changed the function rndn to _rndng10 so that GAUSS 10 and GAUSS 18 would use the same random number generator. (Note that GAUSS 18 uses a higher quality random number generator than GAUSS 10 and _rndng10 should be used only for comparing results between versions.)
With these changes, I get this output
no constant threshold variable = c qn~ns~jm~jn~t0 : 200.0000 200.0000 1.0000 1.0000 4.0000 number of moments: 36.0000 linearity test from analytic : 0.0000 linearity test from averaging: 0.0000 /------ output based on analytic variance formula -----/ proportion of upper regime: 0.1958 estimates and s.e. (threshold~lower regime (lag y,x)~delta (1,lag y,x) 0.3590 0.3875 0.0519 -0.0038 -0.1350 0.4409 -0.0594 -0.0039 0.6580 0.0396 0.0919 0.0387 0.0023 0.0466 0.2601 0.0521 0.0019 0.1322 upper regime estimates and s.e. (lag y,x) 0.8284 -0.0075 -0.0077 0.5229 0.2291 0.0457 0.0033 0.1311 Long-run parameter estimates and s.e. 0.0519 -0.0038 -0.1350 -0.0075 -0.0077 0.5229 0.0696 0.0033 0.1239 1.3404 0.2641 2.3207 Overidentification J-statistic (p-value) 45.0039 0.0162 /------ in case of averaging -----/ proportion of upper regime: 0.1928 estimates and s.e. (threshold~lower regime (lag y,x)~delta (1,lag y,x) 0.3610 0.3862 0.0570 -0.0035 -0.1214 0.4407 -0.0773 -0.0032 0.6044 0.0420 0.0870 0.0407 0.0022 0.0434 0.2432 0.0494 0.0017 0.1207 upper regime estimates and s.e. (lag y,x) 0.8269 -0.0203 -0.0067 0.4830 0.2148 0.0439 0.0031 0.1214 Long-run parameter estimates and s.e. 0.0570 -0.0035 -0.1214 -0.0203 -0.0067 0.4830 0.0701 0.0032 0.1079 1.2555 0.2512 1.9488 Overidentification J-statistic (p-value) 62.2397 0.0001 elapsed time : 11 minutes 27.62 seconds
Based on this, it appears that to reproduce the results of the paper exactly, you would need to use the same starting random seed.
Your Answer
2 Answers
There will be some variation in the output of the code because it uses random numbers in the computations. I ran the code in GAUSS 18 on a Mac and GAUSS 10 on a Linux server and got slightly different results each time.
Then I made two changes which gave me reproducible results.
First change
I set the random number seed, using the rndseed keyword towards the top of the code, like this
new; format /rd /m1 8,4; output file= invest.out on;
tstart = date;
load invest; // Load data and define variables
// Set seed for repeatable random numbers
rndseed 23423;Second change
When running the code with GAUSS 18, I changed the function rndn to _rndng10 so that GAUSS 10 and GAUSS 18 would use the same random number generator. (Note that GAUSS 18 uses a higher quality random number generator than GAUSS 10 and _rndng10 should be used only for comparing results between versions.)
With these changes, I get this output
no constant threshold variable = c qn~ns~jm~jn~t0 : 200.0000 200.0000 1.0000 1.0000 4.0000 number of moments: 36.0000 linearity test from analytic : 0.0000 linearity test from averaging: 0.0000 /------ output based on analytic variance formula -----/ proportion of upper regime: 0.1958 estimates and s.e. (threshold~lower regime (lag y,x)~delta (1,lag y,x) 0.3590 0.3875 0.0519 -0.0038 -0.1350 0.4409 -0.0594 -0.0039 0.6580 0.0396 0.0919 0.0387 0.0023 0.0466 0.2601 0.0521 0.0019 0.1322 upper regime estimates and s.e. (lag y,x) 0.8284 -0.0075 -0.0077 0.5229 0.2291 0.0457 0.0033 0.1311 Long-run parameter estimates and s.e. 0.0519 -0.0038 -0.1350 -0.0075 -0.0077 0.5229 0.0696 0.0033 0.1239 1.3404 0.2641 2.3207 Overidentification J-statistic (p-value) 45.0039 0.0162 /------ in case of averaging -----/ proportion of upper regime: 0.1928 estimates and s.e. (threshold~lower regime (lag y,x)~delta (1,lag y,x) 0.3610 0.3862 0.0570 -0.0035 -0.1214 0.4407 -0.0773 -0.0032 0.6044 0.0420 0.0870 0.0407 0.0022 0.0434 0.2432 0.0494 0.0017 0.1207 upper regime estimates and s.e. (lag y,x) 0.8269 -0.0203 -0.0067 0.4830 0.2148 0.0439 0.0031 0.1214 Long-run parameter estimates and s.e. 0.0570 -0.0035 -0.1214 -0.0203 -0.0067 0.4830 0.0701 0.0032 0.1079 1.2555 0.2512 1.9488 Overidentification J-statistic (p-value) 62.2397 0.0001 elapsed time : 11 minutes 27.62 seconds
Based on this, it appears that to reproduce the results of the paper exactly, you would need to use the same starting random seed.