collapse option in xtabond2

[Sara Harsha]

The model I am working on is dynamic (i.e. lag of dependent variable is one of the regressors). I am running -xtabond2- command in Stata. My supervisor has asked me to use only one lag of instrumental variables. Now, when I am using the option "collapse", the signs of my co-efficients are entirely different than when I am not using it. Can anybody please guide when is the right situation to be using collapse option? Here, I am taking only one lag so should I use the "collapse" option or not?

[Sebastian Kripfganz]

Changing signs when switching from collapsed to standard GMM-style instruments would rather hint towards a more fundamental problem with your data because the underlying assumptions behind these two versions are more or less the same. It is impossible to even speculate about the source of your problem without seeing the command line you typed in Stata and the corresponding regression output.

[Sara Harsha]

Please see the attachment. I have been trying hard to paste Stata output over this screen - couldn't do so. Therefore, I have attached the results.

NIM (net interest margin) is my dependent variable plus, I have many explanatory variables.

Your guidance will be highly appreciated.

[Sebastian Kripfganz]

Before looking at the signs you should first check whether the coefficients are statistically significantly different from zero. Many of them are not. If we cannot reject the hypothesis that the coefficient is different from zero, then the sign does not matter. None of the coefficients that are significant in both version change signs. I would rather ask the question why some of your coefficients are significant in one version but not the other. I suspect that there might be some collinearity issues, meaning that some of your regressors might be highly correlated.

[Sara Harsha]

So many thanks but I am really stuck now. I have no idea how to progress further with my thesis. Two more questions, please:

i) In case some of my regressors turn out to be strongly correlated, what should I do? I can't drop any variable. I want to include all of them in my report.

ii) Please confirm whether the option 'collapse' is required in my case or should I just make do with the standard GMM-syle instrument?

Already grateful!

[Sara Harsha]

Plus, Is there any command other than -estat vce, corr- to measure 'vif' in pane data? Sorry for this bombardment of questions.

[Sebastian Kripfganz]

The number of instruments in the specification with standard GMM-style instruments is indeed (too) large given your relatively small sample. This could in fact explain to some extent the differences in your results. I would not put much confidence in these results but rather trust more the results with the collapsed instruments.

As another comment: Is there any good reason why you supressed the constant term? In general, I would not think that this is a good idea as it could lead to severe bias.

[Sara Harsha]

I have suppressed the constant term because banks cannot generate any profit if there aren't any explanatory variables. Is this wrong thinking? I have just applied a basic intuition. Please confirm if I am wrong.

[Sara Harsha]

There are two variables : 1) Net Advances to Deposit Ratio and II) Advances to Total Assets which are exhibiting a strong correlation. Is it because they have the same numerator? Both of them are extremely important for my research. I can't afford dropping any of them. Is there any way out?

[Sebastian Kripfganz]

I cannot give a good suggestion concerning the choice of your variables because I am not an expert in financial data. That is something you should probably discuss with your thesis supervisor. What I realized is that your variables nadr and eata (those you mentioned?) have almost the same coefficient estimate with and without collapsing and both of them change the sign in the same way. You might ask yourself if these two variables really measure different things. If their variation is just driven by the same underlying variation, say, in "Advances", then you probably should drop one of them. In other words, if they are strongly correlated, you cannot expect different effects anyway. There is just no way to distinguish between them. But again, I do not know the theory behind your model specification.

Concerning the constant term: You can never perfectly control for all the factors that determine profits and your variables will typically also be measured with error. Including the constant term does not really harm. If anything, it might be statistically insignificant which would support your argumentation. Technically, it guarantees that the error term has a zero mean.

[Sara Harsha]

So many thanks to you!

[Manuel Almodovar]

Dr. Kripfganz, let me ask you a question. When you say “Technically, it guarantees that the error term has a zero mean” do you mean just include the error term (regardless of whether it is significant or not) or it must be included and also must be significant? Thank you in advance

[Sebastian Kripfganz]

Whenever a constant term is included, no matter whether it is significant or not, it guarantees that the error term has a mean of zero.

[Manuel Almodovar]

Thank you very much

[Manuel Almodovar]

So, it would be correct to publish a panel with insignificant error term, but the error term should be reported. Is my interpretation correct?