Panel ARDL Bounds steps? I hope you are doing well. I need a little help. Can you please mention the steps for the ARDL Bounds test for cointegration and then causality - the panel data version, not the time series one. If possible, could you also provide a reference as material to study? I've read quite  a few papers and most follow somewhat different steps, different tests and of course, different estimators depending on the data (heterogeneity and/or cross-sectional dependence). I do, however, need to know the steps to the procedure.
Thank you in advance.
 A: Okay, so I found out that the paper "The ARDL Method in the Energy-Growth Nexus Field; Best Implementation Strategies" by A. Menegaki was pretty helpful.
I'm posting this so that people who may have the same questions and problems can see this comment/answer and look it up.
I'll try to summarize the key points here:

*

*Step one would be to check for cross-sectional dependence. The result indicated here, essentially determines the following steps and measures to be taken. B-P LM test, Pesaran's scaled LM and CD tests, as well as Baltagi's corrected scaled LM are some of the go-to choices here.


*Next and importantly, is to check for the order of integration of each variable that is to be used. ARDL can be used w/ I(0) & I(1) variables, but not with I(2), so be careful here. In case of no cross-sectional dependence, there's plenty of Panel Unit Root tests to be used, belonging to the 1st generation of such tests. The researcher should opt for the test, depending on what happens with N,T, homogeneity/ heterogeneity etc. Baltagi's book "Econometric Analysis of Panel Data" has some good stuff for reading, so that they can make an optimal choice given their data. HOWEVER... In the presence of cross-sectional dependence - which is probably the most common thing nowadays, due to globalization, unions etc- it would be best to go for one of the 2nd generation of PUR tests, such as Pesaran's CIPS (2007).


*And we come to cointegration. If c-s dependence is present, then opt for a 2nd gen Cointegration test like Westerlund and Groen-Kleibergen. If not, well, there's a few, like Pedroni, Kao, McCoskey-Kao. Pedroni is the most commonly used, I'd say, given it's 7 (!) stats.


*If cointegration is confirmed, then what's left is to pick an estimator or a few related ones and roll with them, choosing based on the data and slope homogenety tests. And example would be PMG/MG, CCEP/CCEMG (again, in case of c-s dependence). OR, the researcher could use FMOLS/DOLS to get robust long-run estimates. If, on the other hand there's no evidence of cointegration, then, the author says that pooling may be a good idea (Fixed/Random Effects), though I'm not sure why.


*For running causality, Panel Granger & Dumitrescu-Hurlin (Heterogeneity + c-s dependence) are the choices.
EDIT: I think I've found out why pooling may be a good idea at the second to last step; See, according to Phillips & Moon (2000), in a panel data setting, if N is large, spurious regression is not treated like it would be in time series application. If N (and also T, as this is for two-way asymptotics) is large, the relationship is consistently estimated. See also : this question and it's answer
