[Tim Grünebaum]

Thank you Sebastian. That was quick and helpful.
This works quite well and I will keep you updated on results, suggestions and possibly problems :-)

[sonia kaur]

Sebastian Kripfganz Hi sebastain sir...I'm facing a lot of issues in my gmm tests as well....I am doing my thesis on determinants of bank profitability ( unbalanced panel with 500000 obs over 64 quarters) and I have a few explanatory variables. I am unable to know if I have segregated them correctly in my command because I am getting perfect 0.00 values for my ar1 ar2 sargan and Hansen tests...Can you please guide me? My command is as such: xtabond2 roa l.roa s ea em rwa nii llp npl liqa gdpgr cpi fedrate , gmm(l.roa s ea em rwa nii llp npl liqa, collapse) iv(gdpgr cpi fedrate) nolevel

my stata result is as shown:


xtabond2 roa l.roa s ea em rwa nii llp npl liqa gdpgr cpi fedrate , gmm(l.roa s ea em rwa nii llp npl liqa, collapse) iv( gdpgr
> cpi fedrate) nolevel
Favoring speed over space. To switch, type or click on mata: mata set matafavor space, perm.

Dynamic panel-data estimation, one-step difference GMM
------------------------------------------------------------------------------
Group variable: bank Number of obs = 540555
Time variable : quarter Number of groups = 14355
Number of instruments = 569 Obs per group: min = 0
Wald chi2(12) = 252491.94 avg = 37.66
Prob > chi2 = 0.000 max = 62
------------------------------------------------------------------------------
roa | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
roa |
L1. | .1861577 .0013633 136.55 0.000 .1834857 .1888297
|
s | -.0009139 .0000745 -12.26 0.000 -.00106 -.0007678
ea | -.0766093 .0005947 -128.81 0.000 -.077775 -.0754437
em | .1876766 .0004576 410.16 0.000 .1867798 .1885734
rwa | .0042716 .0004254 10.04 0.000 .0034378 .0051053
nii | .0578549 .0006993 82.73 0.000 .0564842 .0592255
llp | -.3106192 .0050583 -61.41 0.000 -.3205333 -.3007052
npl | -.0163417 .0023104 -7.07 0.000 -.0208701 -.0118133
liqa | .0070674 .0005639 12.53 0.000 .0059621 .0081727
gdpgr | .0262086 .0006529 40.14 0.000 .024929 .0274883
cpi | .0001673 .0000158 10.62 0.000 .0001365 .0001982
fedrate | .0002772 8.93e-06 31.03 0.000 .0002597 .0002947
------------------------------------------------------------------------------
Instruments for first differences equation
Standard
D.(gdpgr cpi fedrate)
GMM-type (missing=0, separate instruments for each period unless collapsed)
L(1/63).(L.roa s ea em rwa nii llp npl liqa) collapsed
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z = -227.57 Pr > z = 0.000
Arellano-Bond test for AR(2) in first differences: z = -91.16 Pr > z = 0.000
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(557) = 1.6e+05 Prob > chi2 = 0.000
(Not robust, but not weakened by many instruments.)

Difference-in-Sargan tests of exogeneity of instrument subsets:
iv(gdpgr cpi fedrate)
Sargan test excluding group: chi2(554) = 1.6e+05 Prob > chi2 = 0.000
Difference (null H = exogenous): chi2(3) = 264.14 Prob > chi2 = 0.000

[Sebastian Kripfganz]

Please note that this is primarily a Statalist thread on the xtdpdgmm command, not xtabond2.

A few comments:

• You seem to have a highly unbalanced panel data set with up to 62 observations for some units. Long lags are very likely to be weak instruments. I thus recommend to restrict the maximum lag order of the GMM-type instruments (to probably at most 10).
• Your Arellano-Bond test indicates serial correlation in the idiosyncratic error term. This automatically invalidates the instruments at least for the lagged dependent variable. This could be a consequence of omitted variables. I recommend that you add further lags of the dependent variable and possibly also the independent variables as regressors to the model. Your data set is large enough such that adding further regressors does not pose any problem. You might also consider interaction terms between some of your variables to allow for a richer functional form of the effects. For a discussion, see: Kiviet, J. F. (2019). Microeconometric dynamic panel data methods: Model specification and selection
  issues. MPRA Paper 95159, Munich Personal RePEc Archive.
• Right now, you classify all variables as predetermined. Some of them might be endogenous, in which case the first lag is not a valid instrument. You would need to start with lag 2, assuming there is no serial correlation in the error term any more.
• It is usually recommended to add time dummies to the model.
• Given the highly unbalanced nature of the data, it might be worth using the forward-orthogonal deviations instead of first differences. However, be aware that there is a bug in xtabond2 that could lead to incorrect estimates when using forward-orthogonal deviations. You would need to use the xtdpdgmm command instead.

[sonia kaur]

Sebastian Kripfganz Hi sir, firstly thankyou os os much for replying. I appreciate it so much. Sir can I please trouble you , when you mentioned :

restrict the maximum lag order of the GMM-type instruments (to probably at most 10) and alos this :I recommend that you add further lags of the dependent variable and possibly also the independent variables as regressors to the model. Sir can you please guide me on how to do that? The problem is we were not taught aything on gmm during our masters programme and the professors just told us to add this test in our thesis and i've no clue how to go about with this. I'm so confused and i dont know how to move on.

[Sebastian Kripfganz]

First of all, with xtdpdgmm you can replicate your above model as follows:

xtdpdgmm roa L.roa s ea em rwa nii llp npl liqa gdpgr cpi fedrate, model(diff) collapse gmm(L.roa s ea em rwa nii llp npl liqa, lag(1 .)) iv(gdpgr cpi fedrate, diff) nocons

You can then restrict the maximum lag order for the GMM-type instruments by specifying the suboption lag(1 10). Furthermore, you can add a second lag of the dependent variable and, say, the first lag of the next independent variable s as follows:

xtdpdgmm roa L(1/2).roa L(0/1).s ea em rwa nii llp npl liqa gdpgr cpi fedrate, model(diff) collapse gmm(L.roa s ea em rwa nii llp npl liqa, lag(1 10)) iv(gdpgr cpi fedrate, diff) nocons

(It works similarly with xtabond2.)

[sonia kaur]

Sebastian Kripfganz hi sir, thankyou so much for taking some time out to help me with this. I can't tell you how grateful I am. sir I ran the second command you recommended :

xtdpdgmm roa L

(1/2)

.roa

L(0/1)

.s ea em rwa nii llp npl liqa gdpgr cpi fedrate, model(diff) collapse gmm(L.roa s ea em rwa nii llp npl liqa, lag(1

10

)) iv(gdpgr cpi fedrate, diff) nocons

but my stata is not running it. No results are being formed. Itsomething wrong sir? Something I am missing out?

[sonia kaur]

Sebastian Kripfganz Hi sir, I managed to run this. and this cam eout. I do not have any p values to look at. So how will I know if my result is correct sir?



xtdpdgmm roa L(1/2).roa L(0/1).s ea em rwa nii llp npl liqa gdpgr cpi fedrate, model(diff) collapse gmm(L.roa s ea em rwa nii l
> lp npl liqa, lag(1 10)) iv(gdpgr cpi fedrate, diff) nocons
note: standard errors may not be valid

Generalized method of moments estimation

Fitting full model:
Step 1 f(b) = .00023949

Group variable: bank Number of obs = 540555
Time variable: quarter Number of groups = 14355

Moment conditions: linear = 93 Obs per group: min = 1
nonlinear = 0 avg = 37.65622
total = 93 max = 62

------------------------------------------------------------------------------
roa | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
roa |
L1. | .1238228 .001439 86.05 0.000 .1210024 .1266433
L2. | -.1330753 .0013124 -101.40 0.000 -.1356476 -.1305029
|
s |
--. | -.0056246 .0001765 -31.86 0.000 -.0059706 -.0052786
L1. | .0028136 .0001346 20.91 0.000 .0025498 .0030773
|
ea | -.1206128 .0008507 -141.78 0.000 -.1222801 -.1189455
em | .1717697 .0004793 358.41 0.000 .1708304 .172709
rwa | .0115603 .000518 22.32 0.000 .0105451 .0125756
nii | .0555345 .0008823 62.94 0.000 .0538052 .0572638
llp | -.3433724 .0057227 -60.00 0.000 -.3545886 -.3321561
npl | .1088755 .0033762 32.25 0.000 .1022584 .1154926
liqa | .0081207 .0006465 12.56 0.000 .0068537 .0093878
gdpgr | .0296602 .0009031 32.84 0.000 .0278901 .0314304
cpi | .0000927 .0000169 5.47 0.000 .0000595 .0001259
fedrate | .0005647 .000013 43.51 0.000 .0005393 .0005902
------------------------------------------------------------------------------
Instruments corresponding to the linear moment conditions:
1, model(diff):
L1.L.roa L2.L.roa L3.L.roa L4.L.roa L5.L.roa L6.L.roa L7.L.roa L8.L.roa
L9.L.roa L10.L.roa L1.s L2.s L3.s L4.s L5.s L6.s L7.s L8.s L9.s L10.s L1.ea
L2.ea L3.ea L4.ea L5.ea L6.ea L7.ea L8.ea L9.ea L10.ea L1.em L2.em L3.em
L4.em L5.em L6.em L7.em L8.em L9.em L10.em L1.rwa L2.rwa L3.rwa L4.rwa
L5.rwa L6.rwa L7.rwa L8.rwa L9.rwa L10.rwa L1.nii L2.nii L3.nii L4.nii
L5.nii L6.nii L7.nii L8.nii L9.nii L10.nii L1.llp L2.llp L3.llp L4.llp
L5.llp L6.llp L7.llp L8.llp L9.llp L10.llp L1.npl L2.npl L3.npl L4.npl
L5.npl L6.npl L7.npl L8.npl L9.npl L10.npl L1.liqa L2.liqa L3.liqa L4.liqa
L5.liqa L6.liqa L7.liqa L8.liqa L9.liqa L10.liqa
2, model(diff):
D.gdpgr D.cpi D.fedrate

.

[Sebastian Kripfganz]

After the estimation, you can obtain the Arellano-Bond serial-correlation test by typing estat serial and the Sargan-Hansen overidentification test by typing estat overid. See the help files for details:

help xtdpdgmm
help xtdpdgmm postestimation

[sonia kaur]

Sebastian Kripfganz thankyou so much sebastian. I did manage to run it. But i'm getting this:
Arellano-Bond test for autocorrelation of the first-differenced residuals
H0: no autocorrelation of order 1: z = -174.1354 Prob > |z| = 0.0000
H0: no autocorrelation of order 2: z = 8.6543 Prob > |z| = 0.0000

[sonia kaur]

Sebastian Kripfganz hi Sebastian, I tried changing my command for xtabond2 am i'm getting this currently
xtabond2 roa l2.roa l2.s l2.ea l2.em l2.rwa l2.nii l2.llp l2.npl l2.liqa l.gdpgr cpi fedrate, gmm(l2.roa l2.s l2.ea l2.em l2.rw
> a l2.nii l2.llp l2.npl l2.liqa cpi fedrate, collapse) iv(l.gdpgr) nolevel robust orthogonal nodiffsargan
Favoring speed over space. To switch, type or click on mata: mata set matafavor space, 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: bank Number of obs = 526200
Time variable : quarter Number of groups = 14350
Number of instruments = 675 Obs per group: min = 0
Wald chi2(12) = 11235.47 avg = 36.67
Prob > chi2 = 0.000 max = 61
------------------------------------------------------------------------------
| Robust
roa | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
roa |
L2. | .0837446 .0143128 5.85 0.000 .055692 .1117971
|
s |
L2. | -.0023193 .0001565 -14.82 0.000 -.0026261 -.0020126
|
ea |
L2. | -.0535853 .0021153 -25.33 0.000 -.0577312 -.0494395
|
em |
L2. | -.1377931 .0033978 -40.55 0.000 -.1444526 -.1311335
|
rwa |
L2. | .002203 .0010015 2.20 0.028 .0002401 .0041659
|
nii |
L2. | .0042045 .002908 1.45 0.148 -.001495 .0099041
|
llp |
L2. | -.0519947 .0215567 -2.41 0.016 -.094245 -.0097443
|
npl |
L2. | -.1030135 .0057518 -17.91 0.000 -.1142869 -.0917401
|
liqa |
L2. | -.0008727 .0011141 -0.78 0.433 -.0030563 .001311
|
gdpgr |
L1. | .0334576 .002073 16.14 0.000 .0293947 .0375205
|
cpi | -.0005748 .0000222 -25.90 0.000 -.0006183 -.0005313
fedrate | .0003172 .0000212 14.95 0.000 .0002756 .0003588
------------------------------------------------------------------------------
Instruments for orthogonal deviations equation
Standard
FOD.L.gdpgr
GMM-type (missing=0, separate instruments for each period unless collapsed)
L(1/63).(L2.roa L2.s L2.ea L2.em L2.rwa L2.nii L2.llp L2.npl L2.liqa cpi
fedrate) collapsed
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z = -36.76 Pr > z = 0.000
Arellano-Bond test for AR(2) in first differences: z = 0.93 Pr > z = 0.353
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(663) = 1.4e+05 Prob > chi2 = 0.000
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(663) =10944.79 Prob > chi2 = 0.000
(Robust, but weakened by many instruments.)

[Sebastian Kripfganz]

Just lagging the explanatory variables to avoid endogeneity problems is usually not a good idea. You are ruling out any contemporaneous effect of your right-hand side variables on the dependent variable which is usually not what you want. In particular for the lagged dependent variable, it does not seem to make sense to consider the second lag without the first lag as well.

Finding the right model specification is not a trivial task. I can only refer you again to the working paper by Jan Kiviet that I mentioned in my post #63 above.

[sonia kaur]

Sebastian Kripfganz Hi sebestain..I followed your advice and kpt trying to change my model with econometrics logic and I mange to generate these 3 results...how do they look like to you? I still cant fixmy Hansen and sargan values. is it possible to have this problem if he data is too large?


1ST EXAMPLE
---------+----------------------------------------------------------------
roa |
L1. | .0221422 .0092792 2.39 0.017 .0039552 .0403291
|
s | -.0176074 .0020485 -8.60 0.000 -.0216225 -.0135923
ea | -.5913272 .0668993 -8.84 0.000 -.7224475 -.460207
em | .0886635 .0057107 15.53 0.000 .0774707 .0998562
rwa | .0121 .0067143 1.80 0.072 -.0010597 .0252597
nii | .0534196 .0135918 3.93 0.000 .0267801 .0800591
llp | -.3173799 .3699567 -0.86 0.391 -1.042482 .407722
npl | .2611848 .0756351 3.45 0.001 .1129428 .4094268
liqa | .0029404 .0052084 0.56 0.572 -.0072678 .0131487
stdclroa | 4.017047 .9251678 4.34 0.000 2.203752 5.830343
lz | .074447 .0025463 29.24 0.000 .0694564 .0794376
ne | .0000203 9.43e-06 2.16 0.031 1.86e-06 .0000388
ic | -1.59e-06 9.59e-07 -1.65 0.098 -3.47e-06 2.94e-07
tier2rbcr | 2.003676 .9355492 2.14 0.032 .170033 3.837319
gdpgr | -.0447042 .0092033 -4.86 0.000 -.0627423 -.026666
tvlgdp | -2.77e-06 6.94e-07 -4.00 0.000 -4.14e-06 -1.41e-06
cpi | .0005689 .0000949 6.00 0.000 .000383 .0007548
fedrate | .0005184 .0001031 5.03 0.000 .0003164 .0007204
------------------------------------------------------------------------------
Instruments for orthogonal deviations equation
Standard
FOD.(gdpgr cpi liqa npl rwa fedrate ic tvlgdp)
GMM-type (missing=0, separate instruments for each period unless collapsed)
L2.(L.roa nii)
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z = -1.40 Pr > z = 0.161
Arellano-Bond test for AR(2) in first differences: z = -0.00 Pr > z = 0.997
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(76) = 883.67 Prob > chi2 = 0.000
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(76) = 660.43 Prob > chi2 = 0.000
(Robust, but weakened by many instruments.)



2ND EXAMPLE
----------------------------------------------------------------------------
| Robust
roa | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
roa |
L1. | .0227859 .0092686 2.46 0.014 .0046198 .0409521
|
s | -.017641 .0020664 -8.54 0.000 -.0216911 -.0135909
ea | -.5941819 .0653727 -9.09 0.000 -.7223101 -.4660537
em | .088608 .0057319 15.46 0.000 .0773737 .0998422
rwa | .011926 .0067904 1.76 0.079 -.0013829 .025235
nii | .0516477 .0132199 3.91 0.000 .0257372 .0775581
llp | -.3287297 .3768005 -0.87 0.383 -1.067245 .4097857
npl | .2628361 .0767282 3.43 0.001 .1124516 .4132206
liqa | .0031781 .0052995 0.60 0.549 -.0072088 .013565
stdclroa | 4.042031 .9247331 4.37 0.000 2.229587 5.854474
lz | .0745333 .0025765 28.93 0.000 .0694835 .0795831
ne | .0000206 9.54e-06 2.15 0.031 1.85e-06 .0000393
tier2rbcr | 1.991092 .9436256 2.11 0.035 .1416197 3.840564
gdpgr | -.0440909 .0092653 -4.76 0.000 -.0622506 -.0259313
tvlgdp | -2.83e-06 6.89e-07 -4.10 0.000 -4.18e-06 -1.48e-06
cpi | .0005739 .0000955 6.01 0.000 .0003868 .0007611
fedrate | .0005052 .0001032 4.89 0.000 .0003029 .0007075
------------------------------------------------------------------------------
Instruments for orthogonal deviations equation
Standard
FOD.(gdpgr cpi liqa npl rwa fedrate tvlgdp)
GMM-type (missing=0, separate instruments for each period unless collapsed)
L2.(L.roa nii)
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z = -1.39 Pr > z = 0.166
Arellano-Bond test for AR(2) in first differences: z = 0.02 Pr > z = 0.985
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(76) = 875.88 Prob > chi2 = 0.000
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(76) = 660.45 Prob > chi2 = 0.000
(Robust, but weakened by many instruments.)



3RD EXAMPLE:

| Robust
roa | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
roa |
L1. | .0709386 .0186184 3.81 0.000 .0344471 .1074301
|
s | .1931949 .0860361 2.25 0.025 .0245672 .3618226
s2 | -.0082616 .0035144 -2.35 0.019 -.0151497 -.0013734
ea | -.4951826 .0604164 -8.20 0.000 -.6135965 -.3767686
ic | -2.06e-07 5.52e-07 -0.37 0.709 -1.29e-06 8.75e-07
em | .2072974 .0116413 17.81 0.000 .1844808 .230114
rwa | -.0193173 .0165228 -1.17 0.242 -.0517014 .0130668
rwa2 | .0186785 .0106472 1.75 0.079 -.0021897 .0395466
nii | .0472684 .0113066 4.18 0.000 .0251079 .0694289
llp | -.2983686 .3374817 -0.88 0.377 -.9598206 .3630833
npl | .0087433 .0630474 0.14 0.890 -.1148273 .1323139
liqa | -.0117054 .0045072 -2.60 0.009 -.0205394 -.0028715
stdclroa | 3.799513 .5276648 7.20 0.000 2.765309 4.833717
lz | .0362032 .0039309 9.21 0.000 .0284988 .0439076
ne | .000027 .0000125 2.16 0.030 2.55e-06 .0000515
tier2rbcr | 1.981242 1.485374 1.33 0.182 -.9300373 4.892522
gdpgr | .0117935 .0078676 1.50 0.134 -.0036267 .0272137
tvlgdp | -2.58e-08 1.27e-06 -0.02 0.984 -2.51e-06 2.46e-06
cpi | .0005617 .0000857 6.55 0.000 .0003938 .0007297
fedrate | .0003626 .000127 2.86 0.004 .0001137 .0006115
------------------------------------------------------------------------------
Instruments for orthogonal deviations equation
Standard
FOD.(gdpgr cpi liqa npl rwa ic fedrate tvlgdp rwa2 s2)
GMM-type (missing=0, separate instruments for each period unless collapsed)
L(1/63).(L.roa nii) collapsed
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z = -1.71 Pr > z = 0.088
Arellano-Bond test for AR(2) in first differences: z = 0.57 Pr > z = 0.570
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(115) = 618.16 Prob > chi2 = 0.000
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(115) =1537.79 Prob > chi2 = 0.000
(Robust, but weakened by many instruments.)



What is your opinion on these sir?

[Sebastian Kripfganz]

As mentioned ealier, besides the potential incorrect classification of endogenous regressors as predetermined or exogenous, a likely cause of model misspecification is omitted dynamics, i.e. additional lags of the variables as regressors. Another reason might be that there are further omitted variables that are not in the data set. This would be much harder to fix.

Also note that you would usually want to use the two-step instead of the one-step estimator, in particular when your sample is very large, because the one-step estimator is generally inefficient (unless the quite strong assumption of homoskedasticity is justified). I am afraid that I will probably not be able to help much further because finding the best specification is really data specific. I recommend that you discuss your results and further steps with your thesis advisor.

[sonia kaur]

Sebastian Kripfganz thaankyou so much for all your time and advice. I appreciate your help so much

[Sebastian Kripfganz]

Another update to version 2.2.2 with a few bug fixes is now available on SSC and my personal website:

adoupdate xtdpdgmm, update

You can now also find my presentation slides from this year's London Stata Conference online that provide detailed information on the xtdpdgmm command and the econometrics behind it. If you find this command useful, I would be happy if you could add a reference in your work:

• Kripfganz, S. (2019). Generalized method of moments estimation of linear dynamic panel data models. Proceedings of the 2019 London Stata Conference.