System GMM Hansen Test Problem

[John Vur]

I am running the following command for my research and I just cannot pass the Hansen Test for Overid, which means to accept the null hypothesis.

I simply cannot make the statistic of Hansen Test bigger than 0.05 even though I tried a lot of specifications.

Is there anyone who can advise me on this issues?

Any help would be appreciated very much.

Thank you!


The command and the result are as follows:

xtabond2 log_tfp_op l.log_tfp_op fr log_hhi log_tr log_tr2 i.dbj_securitycode_m if regyear >=2000 & dbj_securitycode_l==1 & ri_total_1977_new>0, gmm(l.log_tfp_op fr, lag(2 5)) iv(log_hhi log_tr log_tr2 ) twostep robust

[Sebastian Kripfganz]

This is an indication of model misspecification. You might want to try adding some further lags of the dependent variable and the regressors to the list of right-hand side variables. Furthermore, I highly recommend to include time dummies.

[Oluwaseun Alo]

Hi Sebastian, I have the same problem. The p values for Hansen test of overidentifying restrictions in virtually all my regressions are 0.00. Since the test of overidentifying restrictions is a test of the coherency of the instruments and my model is correctly specified, can't the Hansen test p values be ignored as in most literatures since AR2 test has shown that there is no autocorrelation in the model? Or maybe I'm getting something wrong, if the p values of Hansen is less than 0.05 level of significance, can't I still accept that the instruments are jointly valid in terms of overidentifying restrictions? though I'm aware of the Roodman rule of thumb of 0.1-0.25. Thanks for shedding more light on this.

[Sebastian Kripfganz]

If the Hansen test is rejected, you cannot infer that the instruments are valid. If you knew already that the instruments were valid, you would not need the Hansen test anymore (but its result would still be at odds with the assumption that they are valid).

Model misspecification cannot only occur if there is remaining serial correlation (indicated by a significant AR(2) test) but also if there are other variables incorrectly classified as predetermined or strictly exogenous when in fact they might be endogenous, or if there are any omitted regressors. Such omitted variables could also be further lags of the dependent variable or the independent variables.

More on GMM estimation of dynamic panel models:

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

[Oluwaseun Alo]

Thank you very much for your response. It's very helpful. I have resolved it now

[Oluwaseun Alo]

Hi Sebastian, I would like to ask if it is appropriate for F-statistics to be too large in GMM. I got a value as high as 679591 in my analysis. Please I would like to ask if such result can be reported. Thank you