I have a panel data where cross section dominates the time dimension i.e. N>T (N around 54 and T around 24). My question is whether its necessary to check for stationarity of the panel data or individual series in this case before i apply Fixed effect/Random Effect and then Arellano Bond. Somewhere i read stationarity is an issue where time dimension dominates cross section which is not in my case.Also, whether arellano bond is an efficient estimator if some of the individual series (specifically independent variables) are of order I(1). Any help will be highly appreciated.
Even if the T is around 24 (i would say pretty good for a time serie) you should check the stationarity. In case you have a dynamic panel RE and FE may be not "so good" estimators. If you have a mix of I(0) , I(1) ARDL aproach is a good option but you have only I(1) also PDOLS/FMOLS could be the right solution. You can check Baltagi, Maddala, Pesaran -anything relating to stationarity and cointegration in panel. There are alot of papers
Rashmi I would second Bogdan. But do bear in mind that if you opt for Internal instrumentation at some point, T>13 may lead to instrument proliferation. I would reckon the correction proposed by Rodman (can't recall the exact name of the paper but it was on institutions and growth). T was reduced to
In STATA there are panel unit root and cointegration tests. Such Im aqnd Pesaran and others. Dynamic Panel System Approach by Blundell and Bond & Arellano and Bouver accounts for stationarity.
Even though you are dealing with a panel dataset where little concern is given in individual time series effects, introducing stochastic trends in your dataset without modeling them would seriously influence the generalization ability of your results.
T=24 may be a pretty good (long) for panel data but I doubt if it T=24 is a sufficient big enough number to yield valid unit root test results (i.e., for time series analysis in such panel). Note that in classical unit root testing (panel or not) your null hypothesis assumes nonstationarity [I(1)] and you test if you should reject this. Unit root tests have been proven to have week power(s) thus it is easy to get a false negative, i.e., the result shows there is no reason to reject the null I(1) even though in fact the process is I(0). There are papers in which authors show this problem based on Monte Carlo experiments. So in a nutshell – yes you should be concerned with your processes in the panel being I(1). But so far I am not familiar with a unit root test that would be able to perform well when T is "rather" small (and T=24 is not a lot in time series analysys), let alone in a panel data framework.
Your panel data dimension (N=54, and T=24) indicates you are working on macro panel data. In macro panel data, stationarity and cross section dependence are issues of concern. So, even if your panel series are (jointly) stationary, the cross-section units is large enough to test for cross-section correlation. If you don't treat the two appropriately, you may likely have estimation with false inference. A test that can address both problems simultaneously is put forward by Pesaran (2006, 2007). All the best
In my opinion, we should also consider economics issues. When time series is small like in the case presented we might not have to worry too much about business cycles, and so on. In any case it is always recommended to calculate a 4 to 5 year average on the time dimension to eliminate both non-stationarity (crude way) and business cycles. In addition, when T=24, stationarity cannot be properly checked (T should be at least 50).
I think that answers of Ileanu and Yusuf are sufficient to respond to the question.
answers of Ileanu and Yusuf are sufficient to respond to the question.
They are as long as you bear in mind that if are unable to reject nonstationarity hipothesis the result is still inconclusive due to to weak power of such test.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Graphs