Two way fixed effects Model , SUR and some problems

[Maria Fleming]

Hello everyone! Hope my post it is not too confusing.

I am studying a Panel Data Model regarding Health and Pensions Expenditures during 17 years in the UE28 countries.
I want to see how the Health and Pensions expenditures determine each other through a SUR process.

But i have some questions:

#1 Should I calculate two different regressions ( two way model in each one, I guess) , and after all this run the SUR with both equations?

#2 I ran Pooled OLS , between , LSVD , Withing and two way fixed effects model , and the significance of the regressors changed sometimes , specially with i introduce i.year to compute the Two-way fixed effects model.. What are the causes of such insignificant in 80% of the variables? And shoul i evaluate the significance directly in the two way fixed effect model?

# 3 How can I correct for heterokedasticity , unit root and serial correlation? How severe is the problem when my sample is n=28 and t=17?

Thank you so much in advanced

[Carlo Lazzaro]

Maria:
welcome to the list.
With n=28 and t=27 there's no valid inference that you can do.
The recipe is, as usual, try your best to collect more data.

[Maria Fleming]

Dear Carlo,

My econometrics professor did not tell me that... She just told me that it is better to use more data. Also , OCDE and other institutions studied panel data models with this kind of data.
Even if i wanted more data, World Bank and Eurostat do not have sufiicient data for almost anythuing regarding health and pensions expenditures and its determinants before 1996.

Kind Regards,

Maria

[Carlo Lazzaro]

Maria:
I've simply foreworded the consequence of having such a scant handful of data before recommending you to use more data.
Your supervisor is seemingly telling you the same: go and get more data (to make valid inference).
Setting apart what above, if you cannot get more data...what lies ahead?

[Maria Fleming]

Dear Carlo,

I will do that , but i will have to select less variables in order to have more years in my data.
Is n=28 enough? Or shoult it be at least 30?

Thank you ,

Maria

[Carlo Lazzaro]

Maria:
i'm not clear with your definition of n.
Do you mean you have n=28 observations per panel?

[Maria Fleming]

Dear Carlo,

I am studying 28 countries ( European Union) during 17 years ( 1996 until 2013) regarding health expenditure and pension expenditure ( public expenditure) .

[Carlo Lazzaro]

Maria:
so you theoretically shoud have 28x17=476 observations (i.e. 17 per each one of the 28 panels included in your analysis).
How many missing values have you detected?

[Maria Fleming]

I am sorry , i made one mistake. It is 18 years , ( 2013-1997 +1). which means i should have 504 observations.
According to STATA commands as xtsum , I have missing values in log of income ( 471), pensions ( 457) , education ( 456) , GDPpc ( 494).

Thank you

[Carlo Lazzaro]

Maria:
tons of missing values, unfortunately.
Does it depend on logging 0 values?
It this is the cause, do not log.
if -log- is not the cause: can't you retrieve missing data or, if unfeasible, -ipolate- or -mi- them (although with more than 90% of missing values I would feel uncomfortable with any approach).

[Nick Cox]

Cross-posted at http://stats.stackexchange.com/quest...d-small-sample

Please note our cross-posting policy, which is that you should tell us about it. http://www.statalist.org/forums/help#crossposting

[Carlo Lazzaro]

Maria.
as an aside to Nick's reminder about cross-posting policy, I noticed differences in your name(s), too across the two forums.
Just for sake of clarity, which one is the right one?

[Maria Fleming]

Dear Carlo , my name is Maria da Alegria Miller Fleming Pinheiro Torres, and sometimes i use the last and other Fleming. But i will correct in the RG forum.

Regarding your comment, the "good" knews is that regardinf public expenditure, none of the variables that have missing values are significant to explain public expenditure on health.
Do you have any suggestion regarding the variables that could explain the expenditures on health and pensions? Besides age structure and GDP?

Thnak you!

Dear Nick i will read the information, i am sorry, I am new in this forum activities!

[Carlo Lazzaro]

Maria:
thanks for your clarifications.
I would recommend you to:
- skimming through the literature of your research field as far as retirement expenditure.is concerned.
As far as the health care expenditure is concerned, I would consider age, # of diseases (coeteris paribus, a high number of chronic comorbidities triggers high level of health care expenditure, at least if the model adopted to deliver health care services is the National Health Service; things might be different for health care systems which are organized differently); deprivation scores; geographical areas (patients living in different part of a given country might total different amounts of health care expenditure, other thinks being equal)
The following link might be useful for more details: http://www.euro.who.int/en/about-us/...m-reviews-hits.
Eventually, a valuable textbook on (panel data) econometrics using Stata is: http://www.stata.com/bookstore/microeconometrics-stata/.

[Maria Fleming]

Thank you so much Carlo!

Last question , to do a sureg i was thinking of :

1. reg healthy x1 x2 ... i.year i.country
2. reg pensions x1 x2 i.year i.country
3. verify if both regressions are "ok" = verify errors, variables significance , etc.
4. If everything is okay with both regressions, i would do the sureg

Am i wrong? Or should i do directly the sureg regression , and verify through that regression the significances ?

Resume of the question : do i have to do two 2-way-fixed effects model first ( for 2 dependent variables) , verify everything as if they were going to be always independent , and finally estimate the SUR?

Thank you again.