cannot run GMM

[Nariman Sayed]

Dear Statalist,


I am using stata v.14... I have unbalanced panel data with T = 17 and N = 18. I mostly have reversal causality from 2 of my control variables (z1 and z2 )and my main independent variable (M1). When I run two-step system GMM, the results change dramatically, eroding the significance of most, if not all, the variables. However, the model is valid according to AR(2), sargan , and Hansan tests, as follows:
Code:

---

xtabond2 lny l.lny lnz1 x1 x2 x3 x4 x5 m1 m1sq m1_z2 z2 x6 x7, ///
gmm (l.lny , lag(1 3) collapse) iv ( lnz1 x1 x2 x3 x4 x5 m1 m1sq m1_z2 z2 x6 x7 ) twostep cluster(country) nodiffsargan



Dynamic panel-data estimation, two-step system GMM
------------------------------------------------------------------------------
Group variable: country                    Number of obs = 177
Time variable : time                          Number of groups = 17
Number of instruments = 17            Obs per group: min = 1
Wald chi2(13) = 1.40e+10               avg = 10.41
Prob > chi2 = 0.000                           max = 16
(Std. Err. adjusted for clustering on country)
------------------------------------------------------------------------------
| Corrected
lny |             Coef.     Std. Err.      z         P>|z|       [95% Conf. Interval]
-------------+----------------------------------------------------------------
lny|
L1. |        1.041343   .0277953  37.46    0.000      .9868654  1.095821
|
lnz1 |      .0005234   .0033601    0.16    0.876    -.0060623 .007109
x1|          .0000164   .0000536    0.31   0.760    -.0000887 .0001214
x2|          .0000259   .0001718    0.15   0.880    -.0003108 .0003625
x3|          .0002729   .0005714    0.48   0.633    -.0008469 .0013927
x4|         -.0000124   .0000428   -0.29   0.772    -.0000962 .0000714
x5|         -.0001993   .0002734   -0.73   0.466    -.0007351 .0003365
m1|         .0052374   .0120197    0.44    0.663    -.0183208 .0287955
m1sq |   -.0018499   .0028094   -0.66      0.510    -.0073563 .0036564
m1_z2 | -.0002276   .0034992   -0.07       0.948    -.0070859 .0066308
z2|         -.0019458   .0062422   -0.31   0.755    -.0141803 .0102886
x6|         -.002183     .0010165   -2.15  0.032    -.0041753 -.0001908
x7|          .0000406   .0000986    0.4    0.680    -.0001525 .0002338
_cons | -.1274088   .1413448    -0.90      0.367    -.4044396 .149622
------------------------------------------------------------------------------
Instruments for first differences equation
Standard
D.(lnz1 x1 x2 x3 x4 x5 m1 m1sq m1_z2 z2 x6 x7)
GMM-type (missing=0, separate instruments for each period unless collapsed)
L(1/3).L.lny collapsed
Instruments for levels equation
Standard
lnz1 x1 x2 x3 x4 x5 m1 m1sq m1_z2 z2 x6 x7
_cons
GMM-type (missing=0, separate instruments for each period unless collapsed)
D.L.lngini collapsed
------------------------------------------------------------------------------
Arellano-Bond test for AR(1) in first differences: z = -3.09 Pr > z = 0.002
Arellano-Bond test for AR(2) in first differences: z = -0.89 Pr > z = 0.375
------------------------------------------------------------------------------
Sargan test of overid. restrictions: chi2(3) = 0.53 Prob > chi2 = 0.913
(Not robust, but not weakened by many instruments.)
Hansen test of overid. restrictions: chi2(3) = 0.49 Prob > chi2 = 0.921
(Robust, but weakened by many instruments.)

---

---

I tried to put the endogenous variables in the gmm style part, however it does not provide much changes in the significance problem. I also tried to change the number of lags, however it is relatively useless as well. So what is the wrong in my command pls.

[Sebastian Kripfganz]

Note that all of the variables you put into the iv() option are assumed to be strictly exogenous not just with respect to the idiosyncratic error term but also the unobserved group-specific effects. This effectively becomes a "random-effects" assumption.

Dynamic panel data GMM estimators are designed for situations in which N is large relative to T. N=18 certainly is very small. This makes it very difficult to obtain reliable estimates for the optimal weighting matrix. Given that asymptotic efficiency cannot be a concern with such a small sample size, I recommend to only use the one-step estimator instead of the two-step estimator. But even then there is no guarantee that you get reliable results with such a small N.

Taking your results at face value, it seems that your dependent variable is very persistent and primarily explained by its own lag. In such a situation, it can be very difficult to find robust predictors even with a good estimator.

[Nariman Sayed]

Many thanks for your response Prof/Mr Sebastian....

- Well, the static model I reached is Fixed effect with very good results in terms of significance of the predictors...

- I wounder about the "random effect " assumption. I do not have observed time invariant variable (e.g gender). Hence, I did not added any such variable as instrument. or am I supposed to add dummies for countries or something like that?

Thanks for your recommendation to try one-step, however it has not improved the results .. I am searching for SEM as well to deal with my problem, although no good initial results I receive so far..

so is there any advice based on your experience can you give in this situation... much appreciated anyway again for your response to this post!