[Neyati Ahuja]

Hello Prof. Sebastian

I have a query in respect of predeterminant variable.
When I took lag of dependent variable as the predetermined variable, with lag (0 0), (0 1),.....(0 .), the resulting coefficient obtained for the lag depenedent variable after running the xtdpdgmm command is positive but not significant. However, when the same is repeated with lag starting from 1 the results for the coefficient is positive as well as significant.
I have checked for instrument validity using Sargan and Hansen Test also.
1. What should be done in such situation.
2. Is there a problem in model specification or we can take lag of predeterminant variable lag starting from 1.
3. If ans to previous question is yes, what should be the starting of lag of endogenous variables. Can it still start from 1.

Regards

[Sebastian Kripfganz]

I am afraid the last xtdpdgmm update (version 2.3.11) was premature and did more harm than good. I "fixed" a bug that actually wasn't one. I apologize for this mishap.

With the now available latest version 2.4.0, the correct computations have been restored for estat serial and estat hausman. Furthermore, a minor bug in option auxiliary has been fixed which was introduced in version 2.3.10.

As a major new feature, this latest version can now compute the continously-updating GMM estimator as an alternative to the two-step and iterated GMM estimators. Simply specify the new option cugmm. The CU-GMM estimator updates the weighting matrix simultaneously with the coefficient estimates while minimizing the objective function. This is in contrast to the iterated GMM estimator (of which the two-step estimator is a special case), which iterates back and forth between updating the coefficient estimates and the weighting matrix. As a technical comment: The CU-GMM objective function generally does not have a unique minimum. The estimator therefore can be sensitive to the choice of initial values. By default, xtdpdgmm uses the two-stage least squares estimates, ignoring any nonlinear moment conditions, as starting values for the numerical CU-GMM optimization. This seems to work fine.

The following example illustrates the CU-GMM estimator, and how the xtdpdgmm results can be replicated with ivreg2 (up to minor differences due to the numerical optimization):

. webuse abdata

. xtdpdgmm L(0/1).n w k, gmm(L.n w k, l(1 4) c m(d)) iv(L.n w k, d) cu nofooter

Generalized method of moments estimation

Fitting full model:

Continously updating:
Iteration 0:   f(b) =  .22189289  
Iteration 1:   f(b) =  .08073713  
Iteration 2:   f(b) =  .07655265  
Iteration 3:   f(b) =  .07646044  
Iteration 4:   f(b) =  .07645679  
Iteration 5:   f(b) =  .07645673  

Group variable: id                           Number of obs         =       891
Time variable: year                          Number of groups      =       140

Moment conditions:     linear =      16      Obs per group:    min =         6
                    nonlinear =       0                        avg =  6.364286
                        total =      16                        max =         8

------------------------------------------------------------------------------
           n | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
           n |
         L1. |   .4342625   .1106959     3.92   0.000     .2173024    .6512225
             |
           w |  -2.153388   .3702817    -5.82   0.000    -2.879126   -1.427649
           k |  -.0054155   .1221615    -0.04   0.965    -.2448477    .2340166
       _cons |   7.284639   1.123693     6.48   0.000     5.082241    9.487037
------------------------------------------------------------------------------

. predict iv*, iv
 1, model(diff):
   L1.L.n L2.L.n L3.L.n L4.L.n L1.w L2.w L3.w L4.w L1.k L2.k L3.k L4.k
 2, model(level):
   D.L.n D.w D.k
 3, model(level):
   _cons

. ivreg2 n (L.n w k = iv*), cue cluster(id) nofooter
Iteration 0:   f(p) =  31.065005  (not concave)
Iteration 1:   f(p) =  27.307398  (not concave)
Iteration 2:   f(p) =  26.543788  (not concave)
Iteration 3:   f(p) =  25.047573  (not concave)
Iteration 4:   f(p) =  24.521102  (not concave)
Iteration 5:   f(p) =  24.107293  (not concave)
Iteration 6:   f(p) =  23.931765  (not concave)
Iteration 7:   f(p) =  23.746613  (not concave)
Iteration 8:   f(p) =  23.636564  
Iteration 9:   f(p) =  23.304181  (not concave)
Iteration 10:  f(p) =  23.241277  (not concave)
Iteration 11:  f(p) =  23.178503  (not concave)
Iteration 12:  f(p) =  23.125314  (not concave)
Iteration 13:  f(p) =  23.074408  
Iteration 14:  f(p) =  19.278726  
Iteration 15:  f(p) =  12.160385  (not concave)
Iteration 16:  f(p) =  11.700402  
Iteration 17:  f(p) =   11.03222  (not concave)
Iteration 18:  f(p) =  10.950583  (not concave)
Iteration 19:  f(p) =  10.907663  
Iteration 20:  f(p) =  10.800048  
Iteration 21:  f(p) =  10.704051  
Iteration 22:  f(p) =  10.703945  
Iteration 23:  f(p) =  10.703942  
Iteration 24:  f(p) =  10.703942  

CUE estimation
--------------

Estimates efficient for arbitrary heteroskedasticity and clustering on id
Statistics robust to heteroskedasticity and clustering on id

Number of clusters (id) =          140                Number of obs =      891
                                                      F(  3,   139) =    83.84
                                                      Prob > F      =   0.0000
Total (centered) SS     =  1601.042507                Centered R2   =   0.5099
Total (uncentered) SS   =  2564.249196                Uncentered R2 =   0.6940
Residual SS             =  784.7107633                Root MSE      =    .9385

------------------------------------------------------------------------------
             |               Robust
           n | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
           n |
         L1. |   .4342987   .1003318     4.33   0.000     .2376521    .6309453
             |
           w |  -2.153233   .2986292    -7.21   0.000    -2.738535    -1.56793
           k |  -.0053816   .1162739    -0.05   0.963    -.2332742    .2225111
       _cons |   7.284114   .8901409     8.18   0.000     5.539469    9.028758
------------------------------------------------------------------------------

To update to the new version, type the following in Stata's command window:

net install xtdpdgmm, from(http://www.kripfganz.de/stata) replace

Disclaimer: I have extensively tested this new version. However, due to the complexity of the command, the variety of options, and the lack of alternative software to compare the results for some advanced options, I cannot guarantee that the implementation is error-free. Please let me know if you spot any irregularities.

[Sebastian Kripfganz]

If either X1 or X2 is endogenous, then it usually makes sense to assume that their interaction X3 is endogenous as well. You can then just treat it the same way as any other endogenous variable.

[Sarah Magd]

############################################
#Interaction variables with xtdpdgmm
#############################################
Can you please write a post about how we should instrument the interaction variable X3 (X1*X2) in the following two cases?
Case 1: X1 and X2 are both endogenous
Case 2: X1 is endogenous whereas X2 is predeterminant

[Sebastian Kripfganz]

In the case of system GMM, yes.

[Sarah Magd]

Thanks, Prof. Sebastian Kripfganz
In this case, do you mean that I should ignore the 1-step weighting matrix chi2 test and report the 2-step weighting matrix chi2 only?

[Sebastian Kripfganz]

As the note says, both 1-step tests are asymptotically invalid because the one-step weighting matrix of the system GMM estimator is not optimal. That is especially true for the first of the two reported tests (1-step moment functions, 1-step weighting matrix). You should just ignore that test.

[Sarah Magd]

Sebastian Kripfganz
Dear Prof. Kripfganz,
I have a question regarding the "estat overid" after running the two-step and one-step system GMM. I run the same specification (i.e., with the same number of lags) and get different results for the overidentification test. For the one-step estimation, I never obtain an insignificant P-value for the 1-step moment functions and 1-step weighting matrix. Even if I change the specification of the model. Please find below the details of the output. My sample has 27 cross-section units and 13 years.
What does this conflicting result mean?

##########################
#estat overid after the one step
##########################
Sargan-Hansen test of the overidentifying restrictions
H0: overidentifying restrictions are valid

1-step moment functions, 1-step weighting matrix chi2(21) = 122.9740
note: * Prob > chi2 = 0.0000

1-step moment functions, 2-step weighting matrix chi2(21) = 27.9882
note: * Prob > chi2 = 0.1405

* asymptotically invalid if the one-step weighting matrix is not optimal
####################################
#estat overid after the two step
####################################
Sargan-Hansen test of the overidentifying restrictions
H0: overidentifying restrictions are valid

2-step moment functions, 2-step weighting matrix chi2(21) = 27.3389
Prob > chi2 = 0.1599

2-step moment functions, 3-step weighting matrix chi2(21) = 27.9939
Prob > chi2 = 0.1403

[Sebastian Kripfganz]

Addendum: For a simple example with industry dummies (not using the two-stage approach), see slide 86 of my 2019 London Stata Conference presentation:

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

[Neyati Ahuja]

Thanks a lot
I ll refer the above paper and will try implement xtseqreg command.

Thank You

[Sebastian Kripfganz]

1. You can include "global" variables (invariant over cross sections) as long as you do not also include time dummies for each year. If those global variables do not account sufficiently for common time effects, then this might result in biased estimates of all coefficients.

2. There is no simple way of analyzing the impact of global variables in the presence of time effects. (In theory, the same two-stage approach as in 3. below could be applied. In practice, due to the necessary correction of the standard errors, this is more complicated because the respective Stata command was not designed to do this for time effects.)

3. In principle, the same issues as above apply to time-invariant variables in the presence of firm-fixed effects. We discuss potential solutions in the following paper:

• Kripfganz, S., and C. Schwarz (2019). Estimation of linear dynamic panel data models with time-invariant regressors. Journal of Applied Econometrics 34 (4), 526-546.

You can add appropriate instruments which are assumed to be uncorrelated with the firm-fixed effects, e.g. by assuming that the time-invariant variables themselves are uncorrelated with the fixed effects. This can be a quite strong assumption. If violated, it could bias all of the coefficient estimates. (With industry dummies, this should be fine. The firm-fixed effects would then need to be interpreted as firm-specific differences which are not due to being in a specific industry. You can then simply add the industry dummy as a regressor and a standard instrument for the level model.) To gain some robustness for the time-varying regressors, you could apply a two-stage procedure (also discussed in the above paper). The latter can be implemented with my xtseqreg command.

[Neyati Ahuja]

Hello Prof. Sebastian

I have a query on variables that can be included in the dynamic panel data model.
In my study I am working on firm level data (N=1400 companies, T=11 years). Besides the firm level variable as independent variables, I am also interested in analysing the impact of macroeconomic variable ((country specific) which are cross section invariant and time variant ) on my dependent variable which is firm performance.
1. Can these time variant cross section invariant variable eg GDP incorportated in the dynamic paniel model along with my time varying cross section varying variable.
2. If ans to 1. is no, is there a way to analyse the impact of these cross section invariant variable.
3. Can we incorporate cross section variant and time invariant variable in the dynamic panel mode eg industry to which the firm belong using a dummy variable. If yes, please elaborate.

Regards

[Sebastian Kripfganz]

Yes, option ue of predict gives you the usual residuals.

[Sarah Magd]

Thanks, Prof Sebastian Kripfganz,

- Yes, when I run two different panel unit root tests, they give contradicting results for the stationarity of ln_GDPc.

- I want to make sure that when I use "predict residual, ue" after estimating the model with xtdpdgmm, this gives the residuals. Am I right?

[Sebastian Kripfganz]

There is an issue at the conceptual level: It does not really make sense to model the effect of a nonstationary variable on a stationary variable (if you believe the stationarity tests). Such a model would be unbalanced in terms of the integration orders. I recommend to do the following:

Case 3: Regress REN on ln_GDPc and L.ln_GDPc. By adding the lag of ln_GDPc, the model in Case 2 becomes nested in Case 3. Case 2 is essentially the same model as Case 3 but with the restriction that the coefficients of ln_GDPc and L.ln_GDPc are identical with opposite signs. If you estimate Case 3 and indeed find that the estimates support this restriction, you can go for the simplified Case 2. Otherwise, I recommend to stick to Case 3.

Case 4: Equivalently, regress REN on ln_GDPc and Δln_GDPc. If the coefficient of ln_GDPc is zero, then again Cases 2, 3, and 4 are equivalent.

With a small time horizon, you do not need to worry about stationarity for the reason it can cause trouble in time series models (nonstandard distributions etc.). However, as mentioned in my previous post, nonstationarity can lead to invalidity of instruments for the level model and weak instruments in general.

Sometimes, it helps to include a deterministic time trend in the regression (and, importantly, in the unit-root tests). Otherwise, the unit-root tests might mistake the deterministic time trend for a unit root.