Re: st: Panel Data Regression - Serial Correlation - Multi-Collinearity - Fixed vs Random Effects

[David Jacobs <jacobs.184@sociology.osu.edu>]

There's much to reply to here.  I can't cover everything.  You 
haven't told us enough so I will have to speculate a bit.

First, although there may be other reasons, the most likely 
explanation for the exclusion of some of your explanatory variables 
is that the variables the fixed-effects routine is dropping remain 
constant over time.  Fixed-effects can't estimate coefficients on 
such variables.  Stata's fixed effects routine typically will not 
drop highly correlated explanatory variables.

Second, you probably need to run fixed-effects if you've done the 
Hauseman test correctly.

Third, the distribution of variables does not affect the degree to 
which they are collinear.  Collinear variables may or may not be 
normal or non normal.  Orthogonal variables (or variables with little 
or no correlation) may or may not be normally  or non normally distributed.

Fourth, you need not employ Spearman correlations if variables are 
not normally distributed.

One quick way to find the amount of serial correlation is to estimate 
an "xtreg" model with ", pa r c(ar1)"  at the end of the command (and 
drop the "fe").  Then after that model has run, type "xtcorr"  The 
last command will give you an estimate of the serial correlations at 
various lags.

Dave Jacobs

At 04:57 PM 1/28/2010, you wrote:
> Hello,
>
> I'm performing a panel data analysis (cross-sectional time series data
> in long format).
>
> Details on my data set:
>
>        panel variable:  id (unbalanced)
>        time variable:  year, 2004 to 2008, but with gaps
>        delta:  1 unit
>
> number of observations is 960 in 358 groups.
>
> One dependent variable, ten explanatory variables (of which 4 are
> industry dummies).
>
> One of my questions is concerning my dummy variables.
>
> After running fe and re regression with xtreg and performing the
> hausman test the result was in favor of the fe regression model.
> Now all my dummy variables are dropped by the regression. Is it save
> to use random effects including the dummies even though the hausman
> test had a p-value below 0.05?
> The LM test for random effects hints at the existence of these
> individual effects.
>
> Question number two is regarding the explanatory variables and their
> possible multi collinearity.
> None of them are normally distributed which is why (if i'm not
> mistaken) i cannot use pwcorr or corr. Instead i could use spearman
> but i have some concerns since i'm
> dealing with possibly auto correlated time-series variables. Is there
> any option for me to get a correlation matrix for my independent
> variables with significance values just like the
> matrix i get when performing pwcorr, sig? I want to use the
> correlation matrix to determine which explanatory variables need to be
> excluded from my regression due to correlation
> above 50%. Also i want to do this pre estimation.
> And more general.. is it useful to look at correlations year-wise or
> on an aggregate level?
>
> The third questions is a follow up to the second. How do i test for
> serial correlation? i used xtserial for each variable. is it save to
> trust or do i have to include all the variables from
> my regression?
>
> Kind Regards Jan
>
> *
> *   For searches and help try:
> *   http://www.stata.com/help.cgi?search
> *   http://www.stata.com/support/statalist/faq
> *   http://www.ats.ucla.edu/stat/stata/


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/