[John Sgr]

Dear Sebastian,

My third question is about the Ahn-Schmidt GMM estimator. I followed your recommendations and underidentification tests show me that I have identification issues with the diff-GMM estimator. I have unbalanced data so I want to use the FOD estimator, does it make sense to add "nl(noserial)" option to the FOD estimator? Or do I need to use model(diff) to use "nl(noserial)" option?

Best regards,
John

[Sebastian Kripfganz]

1. Well, there is no "true" model in applied work. Your test results are quite mixed, so it is hard to make a clear judgement.
2. You can apply the usual tests irrespective of whether the transformed model is in first differences or forward-orthogonal deviations, yes.

[John Sgr]

Dear Sebastian,

Thanks for reply. I re-read the Kiviet (2019) paper and your presentation. It is perfectly clear for me now how to build a model.

In my renewed model Kleibergen-Paap test do not reject both the overidentification (p=0.25) and the underidentification (p=0.27). I am wondering what is the wise step to take now. Should I consider treating some regressors as strictly exogenous as you did on page 112? Also why it took so long to run underidentification test?

Best regards,
John

[John Sgr]

Dear Professor,

To construct maintained statistical model (MSM) I am following your presentation and Kiviet, J. F. (2020). Microeconometric dynamic panel data methods: Model specification and selection issues. Econometrics and Statistics, 13, 16-45. I have several questions about the finalising MSM.

1. Where should I stop to think about alternative MSMs? If the m1,m2,J and incJ are providing good and sufficient results, do I still need to include interaction terms or transform variables? For my model it is hard to interpret results with the interactions. So I do not want to add them while the initial model fulfils the criteria but once I add them MMSC shows improved statistics.
2. For the last two weeks I am updating my model on a daily basis but I need to decide the final model and get the results. What should be the basic expectations from a model to use it as the final model? Sure there would be room for improvement all the time but I am just asking basic requirements to get results.
3. Stage 7 in Kiviet (2020) states that

as long as no problems emerge regarding the coherence tests, impose restrictions on the model by removing regressors with absolute t-ratio’s below 0.5,

. Should I remove all insignificant regressors even if MMSC performance lowers?
4. Is it better to use collapsing and curtailing together? MMSC recommends me the model which only uses collapse rather than both. And the number of instruments is still in the interval that you mention in p.93.

[Sebastian Kripfganz]

1. If there is no good economic/theoretical reason to include these interaction effects and the statistics of the MSM without those interaction terms look reasonably good, I would also not include them. Interpretability of the coefficients in a parsimonious model is also a desirable property.

2. If the m1, m2, and J statistics are acceptable, that could already be sufficient to stop searching for improvements. Keep in mind that any futher improvement might lead to more efficient estimates, but at the cost of less robustness because usually stronger assumptions would be required for each improvement step. Jeff Wooldridge would say: If you assume more, you can gain more. If you assume less, you will lose less.

3. First of all, do not remove any regressor of interest or if theory predicts that a regressor should be relevant. Evidence in your final model that a coefficient is not statistically significant might be quite informative. Besides that, you face a similar tradeoff as before: You could gain efficiency by removing those regressors but at the same time risk misspecification. If in doubt, you could reduce the arbitrariness of your model selection choices by just follow the MMSC.

4. There is no generally accepted convention here. I personally prefer to combine collapsing and curtailing. The MMSc are not suitable to distinguish between collapsing and curtailing because there is no change in the underlying model assumption. If the number of instruments is still reasonable low (probably because you do not have many time periods), then curtailing may not be necessary.

[Annur Wijayakusuma]

Dear Sebastian,

I have a question regarding the lag limit, especially in the xtabond2 command for an endogenous variable. When I used lag limit (0 2) or (1 2) as long as the first lag limit is below 2, the result for AR(1) is significant and AR(2) is rejected which I assumed is a good test result. I also found that the main explanatory variable (RDIntensity) is significant (p<0.005). Second, I run the other command and used a lag limit above 2, for example (2 3) or (2 4) etc, the result for AR(1) is significant and AR(2) is rejected (the same as using lag limit below 2), however, the main explanatory variable (RDIntensity) changed to be insignificant (p>0.005). In this case, which lags limit that I need to choose?

Here I attach the code and the result for you to understand my question

xtabond2 ROA L.RDIntensity L.ROA Leverage L.SIZE i.Year , gmm( RDIntensity , lag(2 2) equation(diff)) gmm( ROA , lag(1 1) equation(level)) gmm( SIZE, lag(1 1) equation(level)) iv( Leverage i.Year, equation (level) )small twostep robust

The result

Dynamic panel-data estimation, two-step system GMM
------------------------------------------------------------------------------
Group variable: ASX_ID                          Number of obs      =      5053
Time variable : Year                            Number of groups   =       460
Number of instruments = 42                      Obs per group: min =         7
F(16, 459)    =     12.48                                      avg =     10.98
Prob > F      =     0.000                                      max =        11
------------------------------------------------------------------------------
             |              Corrected
         ROA |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
 RDIntensity |
         L1. |   .0000643   .0000411     1.56   0.118    -.0000165    .0001451
             |
         ROA |
         L1. |   .1002149   .0482987     2.07   0.039     .0053009    .1951288
             |
    Leverage |  -.0030732   .0014688    -2.09   0.037    -.0059597   -.0001867
             |
        SIZE |
         L1. |   .0047515   .0253567     0.19   0.851    -.0450781    .0545811
             |
        Year |
       2006  |          0  (empty)
       2007  |  -.1203909   .1067071    -1.13   0.260     -.330086    .0893042
       2008  |   -.148882   .1484432    -1.00   0.316    -.4405945    .1428305
       2009  |  -.2447955   .1571492    -1.56   0.120    -.5536167    .0640256
       2010  |  -.1599631   .1495036    -1.07   0.285    -.4537596    .1338333
       2011  |  -.1215046    .160431    -0.76   0.449    -.4367748    .1937657
       2012  |  -.1313113    .171146    -0.77   0.443    -.4676382    .2050156
       2013  |  -.1746189   .1788843    -0.98   0.330    -.5261526    .1769148
       2014  |  -.2223378   .1804264    -1.23   0.218     -.576902    .1322263
       2015  |  -.3130834   .1876878    -1.67   0.096    -.6819174    .0557506
       2016  |     -.2722   .1904794    -1.43   0.154    -.6465199    .1021198
       2017  |  -.1565508   .1957854    -0.80   0.424    -.5412977    .2281961
             |
       _cons |          0  (omitted)
------------------------------------------------------------------------------
Instruments for first differences equation
  GMM-type (missing=0, separate instruments for each period unless collapsed)
    L2.RDIntensity
Instruments for levels equation
  Standard
    Leverage 2006b.Year 2007.Year 2008.Year 2009.Year 2010.Year 2011.Year
    2012.Year 2013.Year 2014.Year 2015.Year 2016.Year 2017.Year
    _cons
  GMM-type (missing=0, separate instruments for each period unless collapsed)
    DL.SIZE
    DL.ROA
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z =  -3.23  Pr > z =  0.001
Arellano-Bond test for AR(2) in first differences: z =  -0.43  Pr > z =  0.669
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(25)   =  54.20  Prob > chi2 =  0.001
  (Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(25)   =  29.78  Prob > chi2 =  0.233
  (Robust, but weakened by many instruments.)

Difference-in-Hansen tests of exogeneity of instrument subsets:
  GMM instruments for levels
    Hansen test excluding group:     chi2(5)    =   9.29  Prob > chi2 =  0.098
    Difference (null H = exogenous): chi2(20)   =  20.49  Prob > chi2 =  0.428
  gmm(RDIntensity, eq(diff) lag(2 2))
    Hansen test excluding group:     chi2(15)   =  20.84  Prob > chi2 =  0.142
    Difference (null H = exogenous): chi2(10)   =   8.94  Prob > chi2 =  0.538
  gmm(ROA, eq(level) lag(1 1))
    Hansen test excluding group:     chi2(15)   =  19.57  Prob > chi2 =  0.189
    Difference (null H = exogenous): chi2(10)   =  10.20  Prob > chi2 =  0.423
  gmm(SIZE, eq(level) lag(1 1))
    Hansen test excluding group:     chi2(15)   =  19.50  Prob > chi2 =  0.192
    Difference (null H = exogenous): chi2(10)   =  10.27  Prob > chi2 =  0.417
  iv(Leverage 2006b.Year 2007.Year 2008.Year 2009.Year 2010.Year 2011.Year 2012.Year 2013.Year 2014.Year 2015.Year 2016.Year 2017.
> Year, eq(level))
    Hansen test excluding group:     chi2(14)   =  20.61  Prob > chi2 =  0.112
    Difference (null H = exogenous): chi2(11)   =   9.17  Prob > chi2 =  0.606

using lag limit below 2

xtabond2 ROA L.RDIntensity L.ROA Leverage L.SIZE i.Year , gmm( RDIntensity , lag(1 2) equation(diff)) gmm( ROA , lag(1 1) equation(level)) gmm( SIZE, lag(1 1) equation(level)) iv( Leverage i.Year, equation (level) )small twostep robust

the result

------------------------------------------------------------------------------
Group variable: ASX_ID                          Number of obs      =      5053
Time variable : Year                            Number of groups   =       460
Number of instruments = 52                      Obs per group: min =         7
F(16, 459)    =     23.17                                      avg =     10.98
Prob > F      =     0.000                                      max =        11
------------------------------------------------------------------------------
             |              Corrected
         ROA |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
 RDIntensity |
         L1. |   .0000456   5.47e-06     8.33   0.000     .0000348    .0000563
             |
         ROA |
         L1. |   .1338519   .0620173     2.16   0.031     .0119789    .2557249
             |
    Leverage |  -.0058227    .002616    -2.23   0.027    -.0109636   -.0006819
             |
        SIZE |
         L1. |   .0060527   .0186599     0.32   0.746    -.0306168    .0427222
             |
        Year |
       2006  |          0  (empty)
       2007  |  -.1151218   .0808547    -1.42   0.155    -.2740131    .0437695
       2008  |  -.1461023   .1146653    -1.27   0.203    -.3714363    .0792317
       2009  |   -.179654   .1332116    -1.35   0.178    -.4414342    .0821261
       2010  |  -.1561344   .1134246    -1.38   0.169    -.3790303    .0667616
       2011  |  -.1201987   .1201033    -1.00   0.317    -.3562192    .1158218
       2012  |  -.1313894   .1273614    -1.03   0.303    -.3816732    .1188944
       2013  |  -.1726875   .1311179    -1.32   0.188    -.4303533    .0849782
       2014  |  -.1956734   .1335396    -1.47   0.144    -.4580982    .0667513
       2015  |  -.2778588   .1557502    -1.78   0.075    -.5839307     .028213
       2016  |  -.3163245   .1474788    -2.14   0.032    -.6061418   -.0265072
       2017  |  -.1556902   .1460305    -1.07   0.287    -.4426614    .1312809
             |
       _cons |          0  (omitted)
------------------------------------------------------------------------------
Instruments for first differences equation
  GMM-type (missing=0, separate instruments for each period unless collapsed)
    L(1/2).RDIntensity
Instruments for levels equation
  Standard
    Leverage 2006b.Year 2007.Year 2008.Year 2009.Year 2010.Year 2011.Year
    2012.Year 2013.Year 2014.Year 2015.Year 2016.Year 2017.Year
    _cons
  GMM-type (missing=0, separate instruments for each period unless collapsed)
    DL.SIZE
    DL.ROA
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z =  -3.23  Pr > z =  0.001
Arellano-Bond test for AR(2) in first differences: z =  -0.50  Pr > z =  0.616
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(35)   =  79.02  Prob > chi2 =  0.000
  (Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(35)   =  44.60  Prob > chi2 =  0.128
  (Robust, but weakened by many instruments.)

Difference-in-Hansen tests of exogeneity of instrument subsets:
  GMM instruments for levels
    Hansen test excluding group:     chi2(15)   =  15.49  Prob > chi2 =  0.417
    Difference (null H = exogenous): chi2(20)   =  29.11  Prob > chi2 =  0.086
  gmm(RDIntensity, eq(diff) lag(1 2))
    Hansen test excluding group:     chi2(15)   =  21.45  Prob > chi2 =  0.123
    Difference (null H = exogenous): chi2(20)   =  23.15  Prob > chi2 =  0.281
  gmm(ROA, eq(level) lag(1 1))
    Hansen test excluding group:     chi2(25)   =  27.48  Prob > chi2 =  0.332
    Difference (null H = exogenous): chi2(10)   =  17.12  Prob > chi2 =  0.072
  gmm(SIZE, eq(level) lag(1 1))
    Hansen test excluding group:     chi2(25)   =  27.15  Prob > chi2 =  0.348
    Difference (null H = exogenous): chi2(10)   =  17.45  Prob > chi2 =  0.065
  iv(Leverage 2006b.Year 2007.Year 2008.Year 2009.Year 2010.Year 2011.Year 2012.Year 2013.Year 2014.Year 2015.Year 2016.Year 2017.
> Year, eq(level))
    Hansen test excluding group:     chi2(24)   =  32.22  Prob > chi2 =  0.122
    Difference (null H = exogenous): chi2(11)   =  12.39  Prob > chi2 =  0.335

[Sebastian Kripfganz]

First of all, you are using an outdated version of xtabond2. You should update it to the latest version because there was a bug concerning the overidentification tests in previous versions. (See the following topic and the follow-up discussion therein: https://www.statalist.org/forums/for...d-xtdpdsys-gmm)

Once you have updated, you then need to check the Hansen test. If the Hansen test rejects the model with lower starting lag order, then you need to use a higher lag. The section on "Model Selection" in my 2019 London Stata Conference presentation might also be useful here:

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

[John Sgr]

Dear Professor,

Should I consider under-identification tests as a part of your point 2?

2. If the m1, m2, and J statistics are acceptable, that could already be sufficient to stop searching for improvements. Keep in mind that any futher improvement might lead to more efficient estimates, but at the cost of less robustness because usually stronger assumptions would be required for each improvement step. Jeff Wooldridge would say: If you assume more, you can gain more. If you assume less, you will lose less.

In the maintained statistical model Kleibergen-Paap test do not reject both the overidentification (p=0.25) and the underidentification (p=0.27). I am wondering what is the wise step to take now. Should I consider treating some regressors as strictly exogenous as you did on page 112? Also is there any way to shorten the duration of running underidentification test?

Best regards,
John

[Annur Wijayakusuma]

Dear Sebastian,

I am just wondering about the Hansen test

Once you have updated, you then need to check the Hansen test. If the Hansen test rejects the model with lower starting lag order, then you need to use a higher lag.

Could you please specify which the Hansen test whether Hansen test of Overidentification or Difference in-Hanse test of exogeneity of the instrument? If it is the difference-in Hansen test, could you specify which one because there are some of the Hansen test results?

[Sebastian Kripfganz]

John Sgr
I am afraid the underidentification test sometimes might take a long time to run. I do not see any workaround. If treating some regressors as strictly exogenous helps with the underidentification tests without compromising on the overidentification tests, that would be a reasonable way to go.

Annur Wijayakusuma
All of these tests suffered from a bug in the previous xtabond2 version. Ideally, neither the overall "Hansen test of overid. restrictions" nor any of the "Difference-in-Hansen tests" should reject the null hypothesis.

[Annur Wijayakusuma]

Dear Sebastian,

I have updated the xtabond2 and read the forum that you share in iterms#22. I tried to run again with gmm, laglimit (1 2), equation (level) and got the result that all the Hansen test were rejected. The thing that confuses me is the endogenous variable need to be instrumented with lag limit 2 and over, however when I run with gmm, laglimit (2 4) and over, equation (diff), the Hansen test accepted null hypotheses. In that case, does it mean that I can use lag limit (1 2) for the endogenous variable as your comment

neither the overall "Hansen test of overid. restrictions" nor any of the "Difference-in-Hansen tests" should reject the null hypothesis.

Here is the command that I applied for laglimit (1 2)

xtabond2 ROA L.RDIntensity L.ROA Leverage L.SIZE Y* , gmm(RDIntensity , lag(1 2) equation(level)) gmm( ROA SIZE, lag(1 1) equation(level)) iv( Leverage Y*, equation (level) )small twostep robust

and the result

Dynamic panel-data estimation, two-step system GMM
------------------------------------------------------------------------------
Group variable: ASX_ID                          Number of obs      =      5053
Time variable : Year                            Number of groups   =       460
Number of instruments = 51                      Obs per group: min =         7
F(17, 459)    =     11.75                                      avg =     10.98
Prob > F      =     0.000                                      max =        11
------------------------------------------------------------------------------
             |              Corrected
         ROA |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
 RDIntensity |
         L1. |   .0000106   1.97e-06     5.37   0.000     6.73e-06    .0000145
             |
         ROA |
         L1. |    .087305   .0419135     2.08   0.038     .0049389    .1696712
             |
    Leverage |  -.0045155   .0015798    -2.86   0.004      -.00762   -.0014109
             |
        SIZE |
         L1. |   .0280836    .030013     0.94   0.350    -.0308964    .0870635
             |
        Year |  -.0002393     .00011    -2.17   0.030    -.0004556   -.0000231
       Y2006 |          0  (omitted)
       Y2007 |    .260118   .1381377     1.88   0.060    -.0113427    .5315787
       Y2008 |   .1969656   .1178522     1.67   0.095    -.0346311    .4285623
       Y2009 |   .0886148   .1231642     0.72   0.472    -.1534208    .3306503
       Y2010 |   .1831819   .1181755     1.55   0.122    -.0490503     .415414
       Y2011 |   .2147642   .1158487     1.85   0.064    -.0128954    .4424239
       Y2012 |   .1909733   .1144421     1.67   0.096     -.033922    .4158686
       Y2013 |   .1218389   .1157598     1.05   0.293     -.105646    .3493237
       Y2014 |   .1126603   .1317305     0.86   0.393    -.1462093      .37153
       Y2015 |          0  (omitted)
       Y2016 |  -.0312178   .1351467    -0.23   0.817    -.2968007    .2343652
       Y2017 |   .1399915   .1357115     1.03   0.303    -.1267014    .4066843
       _cons |          0  (omitted)
------------------------------------------------------------------------------
Instruments for levels equation
  Standard
    Leverage Year Y2006 Y2007 Y2008 Y2009 Y2010 Y2011 Y2012 Y2013 Y2014 Y2015
    Y2016 Y2017
    _cons
  GMM-type (missing=0, separate instruments for each period unless collapsed)
    DL.(ROA SIZE)
    DL(1/2).RDIntensity
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z =  -3.27  Pr > z =  0.001
Arellano-Bond test for AR(2) in first differences: z =  -0.79  Pr > z =  0.427
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(33)   =  85.56  Prob > chi2 =  0.000
  (Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(33)   =  38.74  Prob > chi2 =  0.226
  (Robust, but weakened by many instruments.)

Difference-in-Hansen tests of exogeneity of instrument subsets:
  gmm(RDIntensity, eq(level) lag(1 2))
    Hansen test excluding group:     chi2(14)   =  21.31  Prob > chi2 =  0.094
    Difference (null H = exogenous): chi2(19)   =  17.43  Prob > chi2 =  0.561
  gmm(ROA SIZE, eq(level) lag(1 1))
    Hansen test excluding group:     chi2(13)   =  10.60  Prob > chi2 =  0.644
    Difference (null H = exogenous): chi2(20)   =  28.14  Prob > chi2 =  0.106
  iv(Leverage Year Y2006 Y2007 Y2008 Y2009 Y2010 Y2011 Y2012 Y2013 Y2014 Y2015 Y2016 Y2017, eq(level))
    Hansen test excluding group:     chi2(21)   =  28.58  Prob > chi2 =  0.124
    Difference (null H = exogenous): chi2(12)   =  10.16  Prob > chi2 =  0.602

When I applied laglimit (2 4and over ) equation (diff) , here is the command and result

xtabond2 ROA L.RDIntensity L.ROA Leverage L.SIZE Y* , gmm(RDIntensity , lag(2 4) equation(diff)) gmm( ROA SIZE, lag(1 1) equation(level)) iv( Leverage Y*, equation (level) )small twostep robust

------------------------------------------------------------------------------
Group variable: ASX_ID                          Number of obs      =      5053
Time variable : Year                            Number of groups   =       460
Number of instruments = 59                      Obs per group: min =         7
F(17, 459)    =      7.64                                      avg =     10.98
Prob > F      =     0.000                                      max =        11
------------------------------------------------------------------------------
             |              Corrected
         ROA |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
 RDIntensity |
         L1. |   .0000869   .0001078     0.81   0.421    -.0001249    .0002987
             |
         ROA |
         L1. |   .1672756   .0624908     2.68   0.008     .0444721    .2900792
             |
    Leverage |  -.0057871    .002646    -2.19   0.029    -.0109869   -.0005872
             |
        SIZE |
         L1. |   .0131653   .0304949     0.43   0.666    -.0467616    .0730922
             |
        Year |  -.0002008     .00012    -1.67   0.095    -.0004365    .0000349
       Y2006 |          0  (omitted)
       Y2007 |   .2712584   .1411832     1.92   0.055    -.0061871    .5487039
       Y2008 |   .2306392   .1112218     2.07   0.039     .0120721    .4492063
       Y2009 |   .1739117   .1168449     1.49   0.137    -.0557055    .4035289
       Y2010 |   .2316094   .1086062     2.13   0.033     .0181825    .4450364
       Y2011 |   .2525564   .1047596     2.41   0.016     .0466884    .4584243
       Y2012 |   .2207063   .0989822     2.23   0.026     .0261919    .4152208
       Y2013 |   .1743497   .0957391     1.82   0.069    -.0137915    .3624909
       Y2014 |   .1532922   .1018045     1.51   0.133    -.0467685    .3533528
       Y2015 |          0  (omitted)
       Y2016 |  -.0138042   .1147283    -0.12   0.904     -.239262    .2116535
       Y2017 |   .1889721   .1190594     1.59   0.113    -.0449969    .4229412
       _cons |          0  (omitted)
------------------------------------------------------------------------------
Instruments for first differences equation
  GMM-type (missing=0, separate instruments for each period unless collapsed)
    L(2/4).RDIntensity
Instruments for levels equation
  Standard
    Leverage Year Y2006 Y2007 Y2008 Y2009 Y2010 Y2011 Y2012 Y2013 Y2014 Y2015
    Y2016 Y2017
    _cons
  GMM-type (missing=0, separate instruments for each period unless collapsed)
    DL.(ROA SIZE)
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z =  -3.21  Pr > z =  0.001
Arellano-Bond test for AR(2) in first differences: z =  -0.26  Pr > z =  0.798
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(41)   = 126.92  Prob > chi2 =  0.000
  (Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(41)   =  58.75  Prob > chi2 =  0.036
  (Robust, but weakened by many instruments.)

Difference-in-Hansen tests of exogeneity of instrument subsets:
  GMM instruments for levels
    Hansen test excluding group:     chi2(21)   =  21.86  Prob > chi2 =  0.408
    Difference (null H = exogenous): chi2(20)   =  36.90  Prob > chi2 =  0.012
  gmm(RDIntensity, eq(diff) lag(2 4))
    Hansen test excluding group:     chi2(14)   =  20.18  Prob > chi2 =  0.125
    Difference (null H = exogenous): chi2(27)   =  38.57  Prob > chi2 =  0.069
  gmm(ROA SIZE, eq(level) lag(1 1))
    Hansen test excluding group:     chi2(21)   =  21.86  Prob > chi2 =  0.408
    Difference (null H = exogenous): chi2(20)   =  36.90  Prob > chi2 =  0.012
  iv(Leverage Year Y2006 Y2007 Y2008 Y2009 Y2010 Y2011 Y2012 Y2013 Y2014 Y2015 Y2016 Y2017, eq(level))
    Hansen test excluding group:     chi2(29)   =  37.12  Prob > chi2 =  0.143
    Difference (null H = exogenous): chi2(12)   =  21.64  Prob > chi2 =  0.042

I look forward to your review

Regards,

Annur Wijayakusuma

[Sebastian Kripfganz]

Neither of your examples uses valid instruments. Untransformed lags of the dependent variable ROA cannot be used as instruments for the level model because they are correlated by construction with the unobserved unit-specific "fixed effects". Similar arguments apply to the other variables if they are allowed to be correlated with those unobserved effects. Please consult my 2019 London Stata Conference presentation for details and examples on how to specify instruments.

Furthermore, it seems like you are still using an old version of xtabond2. In the latest version, the omitted time dummies would no longer be displayed in the regression output. The consequence of the bug are incorrect degrees of freedom for the overidentification tests. For example, the Hansen test in your first example should have 36 degrees of freedom, not 33. Thus, also the p-values are incorrect.

[Annur Wijayakusuma]

Dear Sebastian,

Please correct me if I am wrong about your comment below:

Untransformed lags of the dependent variable ROA cannot be used as instruments for the level model because they are correlated by construction with the unobserved unit-specific "fixed effects". Similar arguments apply to the other variables if they are allowed to be correlated with those unobserved effects

based on your presentation in London page 31
1. lagged dependent variable (ROA) cannot be used as an instrument if I used lag limit (1 1) because of correlation with unobserved error. I need to use lag 2 and over of ROA to be a valid instrument
2. RDIntensity which is an endogenous variable, I need to lag to 2 and over.

What about the SIZE variable that I lagged in the regression, can I use lag (1 1) as this assumed strictly exogenous or I need to lag (2 2) or over?

it seems like you are still using an old version of xtabond2

I updated the xtabond2 by using command:

ssc install xtabond2, replace

based on https://www.statalist.org/forums/for...to-do-xtdpdgmm. Am I doing right?

Regards,

Annur Wijayakusuma

[Sebastian Kripfganz]

Apologies. I forgot that xtabond2 by default applies a first-difference transformation to the GMM-style instruments for the level equation. In this case, the first lag of the (first-differenced) dependent variable is a valid instrument for the level equation, provided there is no serial correlation in the idiosyncratic error term. The same is true for any other predetermined or endogenous variable. For the equation in first differences, you would need to start with the second lag, however.

The way you were updating the command should usually work, yes.