[Gretl-users] A Newbie Needs Help

[Sven Schreiber]

Am 10.03.2008 23:57, Scott David Orr schrieb:
> I took a class in causal modeling more than 10 years ago, and while I 
> thought I remembered the basics, since then all my work has been with 
> structural equation models, and I find I'm now a bit lost....

You have my sympathy and understanding, but I doubt that there's any 
quick solution to your problem...

> 
> Let me explain what I'm trying.  Basically, I'm trying to test the 
> hypothesis that high levels of press freedom tend to prevent violent 
> ethnic conflict, because ethnic groups can fight things out in the 
> media.  Therefore, the main effect I'm looking for is an effect of media 
> freedom and ethnic violence, and my guess is that effect will be a bit 
> lagged, though I'm not sure of that, and it's also possible that each 
> variable affects the other.  I have data at least back to 1990 in many, 
> many countries for both of these, though I intend to do the tests just 
> in sub-Saharan Africa and post-Communist Europe.
> 
> Other endogenous variables that could affect the equation would be 
> democracy (the Freedom House political freedom score), unemployment, and 
> change in per-capita GDP.  I'm working on figuring out exogenous 
> variables, but election years and possibly the presence of droughts look 
> good, and literacy rates (separately for men and women) might also be 
> useful.
> 
> My question is, how do I frame this.  Basically, I should have time 
> series data for each variable for each of the countries in question.  
> Each country could therefore be analyzed individually, but I'd ideally 
> expect patterns within particular regions, if not across regions.  My 
> memory vaguely recalls that I want to use SURE or some kind of 
> simultaneous equations analysis, but I've been looking through the two 
> relevant texts I have (Gujarati, Third Edition, and Hamilton's Time 
> Series Analysis), and come to the conclusion that I'm a lot less smart 
> than I thought I was, at least on this subject.

The question is if you're ready to assume and then exploit some degree 
of homogeneity (equal parameter values) across countries. If so, you're 
in a panel context. If not, then you could use SURE. Country-per-country 
is also admissible, it's all a matter of efficiency and sample size.

The bigger problem that I see is your set of endogenous explanatory 
variables, so you may have to use some instrumental-variables approach.

> 
> Could anyone give me a few pointers?  And if those pointers included 
> tips on setting this up in GRETL, that would also help. 

Putting all the ingredients together is definitely doable but is a 
full-fledged research project I'd say. As I said, I don't think there's 
a quick solution.

  One specific
> question I have what do to with exogenous variables that don't vary much 
> over time.  To wit, I'm suspect literacy rates play a role, but since 
> they don't change much over time, that roles should be seen across 
> countries rather than over time within countries (which is one reason a 
> multiple-country analysis would be useful).

Yes then you need a panel analysis. However, those time-constant 
variables are hard (if not impossible) to distinguish from (other) fixed 
effects. So you would have to hope you don't need to use a fixed-effects 
model.


cheers,
sven