I have a data set from almost 20000 companies, from 8 years.
and it is an unbalanced data set.
I searched for a flowchart about all steps I have to do to analyze a panel like this.but, unfortunately, I couldn't find a nice conclusion.
but this is what I realize till now:
1- we should do a unit root test for all variables.if they were stationary or
Cointegrated then we can use OLS
2-we check if it is pooling data or panel data.
3-then Husman test to decide between fixed or random effect method
4- creating model
5- doing Waldrich test for autocorrelation and heteroscedastic (likelihood ratio)
6- check if error term is normal,if not find a way to fix it
now my questions are:
1-are my steps right?should i add something or not
(for example i saw on net someone said with little T,first step is not necessary and if it was unit root still follow these steps!)
2- variable selection in panel data
.i use R,and i realized step()doesn't work for panel data.I try ti logically eliminate some variables, now can I just use the backward method?is it efficient?
3- in my panel according to tests fixed model was a good model, but in results, the between model looked much better!
is it because I didn't take the variable selection part so serious yet? or it is possible that tests show we have individual effects but still between method works better?
4- since my dependent variable is cost, and my data is in 8 years, how should I consider the inflation rate? (should I just ignore cause cost variable was stationary according to tests)
thanks for your attention.
Best Answer
As I check out your steps. I have never seen all those steps straight forward in one place. So, well done. it's a good summary. Just, there is something about the stationary test. If the panel doesn't include many years (not a long panel), you don't need to do that test.