Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
From | Federica Di Marcantonio <dmfederica@yahoo.it> |
To | statalist@hsphsun2.harvard.edu |
Subject | st: Panel model specification |
Date | Thu, 14 Nov 2013 20:46:10 +0100 |
Dear Statalisters, I have a balanced panel (41 SSA countries from 1968 to 2008). In order to check all the issues relate to panel data I proceded as follow: run ols with id and year dummies and then check if they are significant noi xi: reg logfood_pin2004 logland logirrigation logfertilizer_fao loganimtrans logtractors logurb loglabour logexp logcereal_aid logbatt landl raindu p1 p3 i.id i.year noi testparm _Iid* noi testparm _Iyear* (significant so I include them in the FE) Then I generate a time trend ( I don't remember where, but I read that time trend has to be retained if significant but I am not sure why it is so, can you explain and tell me if I have to retain it?) g t=year-1968 noi tab id, g(ctry) it is significant but then I have also generate COUNTRIES TREND forvalues i=1(1)40{ gen tr`i'=t*ctry`i' } I remember to read that if the trends are significantly different across states and cannot be combined into a single overall keep the country trend. In my case considering that the trend is significant I have to keep only t So with this model I have runned the FE and RE model xi: xtreg logfood_pin2004 logland logirrigation logfertilizer_fao loganimtrans logtractors logurb loglabour logexp logcereal_aid logbatt landl raindu p1 p3 t i.year, fe then I tested for cross-sectional dependence - Pasaran CD (xtcsd, pesaran abs), for heteroskedasticity (xttest3), and serial correlation...all the problems are present in my model so I corrected them by using xtpcse, corr(ar1) I also tested for unit root and variables are stationary (some with drift some with trend). However, I also did a second checking for unit root by calculating the residual of my final model (pcse) and test the residual for unit root (predict e, resid; xtfisher e ). Here the results Fisher Test for panel unit root using an augmented Dickey-Fuller test (0 lags) Ho: unit root chi2(74) = 219.5621 Prob > chi2 = 0.0000 Nevertheless, since I wanted to include time invariant var i did a Three stages FEVD and in the last stage i used pcse My questions are: 1) do I have to retain the trend? 2) Is my procedure correct? 3) what can I do in order of understanding if my estimator fit well the model? Thanks Federica Di Marcantonio * * For searches and help try: * http://www.stata.com/help.cgi?search * http://www.stata.com/support/faqs/resources/statalist-faq/ * http://www.ats.ucla.edu/stat/stata/