Problem with constant and time dummies in XTABOND2

[George Kladakis]

Hello,

I'm working with system GMM in the xtabond2 command.

However, I'm having a problem with the constant. In most of my regressions, the constant is omitted. I'm using both time and country dummies. The constant is never omitted when I remove the time dummies. My data contains a global panel of banks which is unbalanced but without gaps within banks.

Here is how one of my commands looks like:

xtabond2 lc1 L.lc1 wstrat pl roaa lnta eqrat mq rgdpg unemp infg i.year i.countryidw1 , gmm( L.( wstrat pl lc1 roaa lnta eqrat mq ), laglimit(1 3) ) iv(i.year i.countryidw1 rgdpg unemp infg) robust twostep

In which cases it is ok to suppress the constant with noconstant? How could I run the regressions without the omission of the constant?

I would appreciate any help!

[Sebastian Kripfganz]

This is due to collinearity of the intercept with the set of time dummies (and possibly also the set of country dummies). You have included too many of those dummies and need to omit some (at least one) of them. Note that there is also a bug in xtabond2 that could produce incorrect degrees of freedoms for the Sargan/Hansen overidentification tests (and therefore incorrect p-values) if there are some omitted coefficients in the output, which generally happens when you include time dummies.

To avoid this problem, you might want to consider using my xtdpdgmm command as an alternative to xtabond2. The xtdpdgmm command also has the option teffects that automatically adds the correct number of time dummies.

More information on the xtdpdgmm command, the above-mentioned problem with time dummies, and dynamic panel data GMM estimation in Stata in general:

• Kripfganz, S. (2019). Generalized method of moments estimation of linear dynamic panel data models. Proceedings of the 2019 London Stata Conference.
• XTDPDGMM: new Stata command for GMM estimation of linear (dynamic) panel data models

[George Kladakis]

Thank you very much Sebastian. Indeed, after dropping the last year dummy, the constant is always there. I will certainly give xtdpdgmm a try!

[Khairul Nizam Osman]

Hi Sebation, My name is Khairul. Hopefully, my issue is related to this post. I'd really appreciate if you could help me to explain why the coefficient indep var "garch" is omitted after adding y*(year dummy), and how to correct it in GMM. Thanks in advance.



xtabond2 lev1_n l.lev1_n l.garch l.tang_n l.prof_n l.size_n l.MALSMCTG y* if manu_dummy==1 , gmm (l.lev1_n l.tang_n l.prof_n l.size_n) iv(l.garch l.MALSMCTG y*)
> nolevel robust nodiffsargan
Favoring space over speed. To switch, type or click on mata: mata set matafavor speed, perm.
Warning: Two-step estimated covariance matrix of moments is singular.
Using a generalized inverse to calculate robust weighting matrix for Hansen test.

Dynamic panel-data estimation, one-step difference GMM

Group variable: code Number of obs = 1209
Time variable : YEAR Number of groups = 186
Number of instruments = 119 Obs per group: min = 1
Wald chi2(15) = 115.88 avg = 6.50
Prob > chi2 = 0.000 max = 7

Robust
lev1_n Coef. Std. Err. z P>z [95% Conf. Interval]

lev1_n
L1. .5269037 .099638 5.29 0.000 .3316168 .7221906

garch
L1. 0 (omitted)

tang_n
L1. .0708908 .0502467 1.41 0.158 -.0275908 .1693725

prof_n
L1. -.0693423 .0462812 -1.50 0.134 -.1600517 .0213672

size_n
L1. .0086365 .0457956 0.19 0.850 -.0811213 .0983943

MALSMCTG
L1. .0005266 .0001928 2.73 0.006 .0001487 .0009045

y_2010 0 (omitted)
y_2011 0 (omitted)
y_2012 .0079653 .0053868 1.48 0.139 -.0025927 .0185232
y_2013 -.0069351 .0045478 -1.52 0.127 -.0158485 .0019784
y_2014 -.0066782 .0044944 -1.49 0.137 -.0154871 .0021308
y_2015 -.0089414 .0041891 -2.13 0.033 -.0171518 -.000731
y_2016 -.003501 .0037236 -0.94 0.347 -.0107992 .0037972
y_2017 0 (omitted)
y_2018 -.0042253 .0059673 -0.71 0.479 -.0159209 .0074703

Instruments for first differences equation
Standard
D.(L.garch L.MALSMCTG y_2010 y_2011 y_2012 y_2013 y_2014 y_2015 y_2016
y_2017 y_2018)
GMM-type (missing=0, separate instruments for each period unless collapsed)
L(1/8).(L.lev1_n L.tang_n L.prof_n L.size_n)

Arellano-Bond test for AR(1) in first differences: z = -3.92 Pr > z = 0.000
Arellano-Bond test for AR(2) in first differences: z = -1.54 Pr > z = 0.122

Sargan test of overid. restrictions: chi2(104) = 166.06 Prob > chi2 = 0.000
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(104) = 111.47 Prob > chi2 = 0.290
(Robust, but weakened by many instruments.)

[Sebastian Kripfganz]

There is apparently perfect collinearity between garch and some of the other variables, or garch may be constant over time.

[Khairul Nizam Osman]

Noted with thanks, Sebastian. In this case, would it be okay for not adding y*? Apparently, without y*, the "garch" coefficient is not omitted, but I would mention in my paper why I choose a model without y*. Again, thank you.

[Sebastian Kripfganz]

Given that your time effects are mostly not statistically significant, this would be an option. You could also just add y_2014-y_2018 instead of y*, but then the coefficient of garch might still just capture a time base effect.

[Gabriel Temesgen]

hello Sebastian Kripfganz I want to use Panel VAR with two-step GMM model. how can I estimate the impulse response functions? Is the XTABOND2 command works with panel VAR model?
Thanks