Re: st: autocorrelation in panel data

[Rufus Peabody <rufus.peabody@gmail.com>]

To make this more concrete, here is what I have done--as you can see,  
the autocorrelation is definitely the further back you lag, but there  
is a lot of noise in this that I'd like to eliminate.  I.e. in the  
future I would not expect kpct to be more correlated with L4.kpct than  
with L1.kpct.

. regress kpct LY_adjkpct_norm LY_adjkpct_norm_Iobs L(1/10).kpct if  
L10.year==year

      Source |       SS                df             MS               
			Number of obs =   18923
-------------+------------------------------          					 F( 12,  
18910) =  357.69
       Model |  28.7720099       12     2.39766749          		 Prob >  
F      =  0.0000
 Residual |  126.758615     18910  .006703258         		  R- 
squared     =  0.1850
-------------+------------------------------           					Adj R- 
squared =  0.1845
       Total |  155.530625      18922    .008219566          		 Root  
MSE      =  .08187

------------------------------------------------------------------------------
        kpct |      Coef.   Std. Err.      t    P>|t|     [95% Conf.  
Interval]
------------- 
+----------------------------------------------------------------
LY_adjkpct~m |   .0423668   .0031874    13.29   0.000     .0361193    . 
0486143
LY_adjkpct~s |   9.46e-06   2.15e-06     4.40   0.000     5.25e-06    . 
0000137
        kpct |
         L1. |   .0670691   .0072904     9.20   0.000     . 
0527794    .0813589
         L2. |   .0611926   .0073193     8.36   0.000     . 
0468461    .0755391
         L3. |   .0634665   .0073307     8.66   0.000     . 
0490977    .0778354
         L4. |   .0710584   .0073266     9.70   0.000     . 
0566976    .0854192
         L5. |    .067266   .0073745     9.12   0.000     . 
0528113    .0817207
         L6. |   .0568484    .007405     7.68   0.000      . 
042334    .0713628
         L7. |   .0420981   .0073926     5.69   0.000     . 
0276079    .0565882
         L8. |   .0485835   .0073626     6.60   0.000     . 
0341522    .0630149
         L9. |   .0450147    .007324     6.15   0.000     . 
0306591    .0593704
        L10. |   .0513367   .0072931     7.04   0.000     . 
0370416    .0656318
       _cons |   .0209256    .002381     8.79   0.000     . 
0162587    .0255925
------------------------------------------------------------------------------




On Aug 19, 2008, at 12:54 AM, Rufus Peabody wrote:


> Hi all,
>
> I'm dealing with data where I'm trying to predict a variable based  
> on previous values of that variable.  I have an unbalanced panel.  I  
> have no problems getting the coefficient for each lag of the variable:
>
> (data is already tsset with a panel variable and time variable)
>
> foreach v of local vars {
> 	regress `v' L(1/10).`v' LY_`v'_norm LY_`v'_norm_intobs if  
> L10.year==year
> }
>
> , where `v' is the variable, LY_`v'_norm is the normalized average  
> of the variable for the particular panel during the previous year &  
> LY_`v'_norm_intobs is an interaction of LY_`v'_norm and the number  
> of observations for that panel the previous year.  (Since there is a  
> lot of luck inherent in the variable in a short-run period, I'm  
> including these so I can "regress to the mean" the variable as  
> appropriate.
>
> While the results of this regression are not problematic (shown  
> below), I would like a way to fit a curve which takes out all the  
> noise from including a bunch of lags.   - corrgram - does not seem  
> to work with panel data.  I should point out that there are no fixed  
> effects.  I would like to fit a decay function that I can apply to  
> all panels.
>
> Thanks!
> -Rufus
> *
> *   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/