st: rollreg question

[Cameron Hooper <chooper@umich.edu>]

Hello

My question relates to the 'rollreg' routine available via ssc install.

I have panel data on which I wish to perform rolling regressions. 
However, for consistency with prior research I want to use standard OLS 
regressions as produced by "regress" rather than ivreg (or ivreg2). Here 
is an example of problem.


. use http://www-personal.umich.edu/~chooper/stata/rrtest1, clear
. list

     +----------------------------+
     | id    y    x1    x2   year |
     |----------------------------|
  1. |  1   12    80   100   1980 |
  2. |  1   23    90    93   1981 |
  3. |  1   45   100    62   1982 |
  4. |  1   13    56   120   1983 |
  5. |  1   22    65    90   1984 |
     |----------------------------|
  6. |  1   21    72    80   1985 |
  7. |  1   56   120    12   1986 |
  8. |  1   43    90    45   1987 |
  9. |  1   29    67    67   1988 |
 10. |  1   14    42    55   1989 |
     |----------------------------|
 11. |  2   34    76    87   1985 |
 12. |  2   56   123    34   1986 |
 13. |  2   32    56    56   1987 |
 14. |  2   67    90    98   1988 |
 15. |  2   83    54    54   1989 |
     |----------------------------|
 16. |  2   45    67    76   1990 |
 17. |  2   12    43    23   1991 |
 18. |  2   45    56    56   1992 |
 19. |  2   23    43    67   1993 |
 20. |  2   56    56   120   1994 |
     +----------------------------+

// First I run a "standard" OLS. I will want to obtain the same results
// when I run rollreg.

. regress y x1 x2 if id == 1 & year < 1985

      Source |       SS       df       MS       Number of obs =       5
-------------+------------------------------    F(  2,     2) =    5.88
       Model |  603.433889     2  301.716944    Prob > F      =  0.1453
    Residual |  102.566111     2  51.2830556    R-squared     =  0.8547
-------------+------------------------------    Adj R-squared =  0.7094
       Total |         706     4       176.5    Root MSE      =  7.1612

------------------------------------------------------------------------------
   y |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
   x1 |  -.0532435   .3456986    -0.15   0.892    -1.540665    1.434178
   x2 |  -.6242087   .2963772    -2.11   0.170    -1.899417    .6509993 
_cons |   85.21505   52.12469     1.63   0.244    -139.0594    309.4895
------------------------------------------------------------------------------

// Now I try to reproduce this using rollreg

. rollreg y x1 x2 if id == 1, move(5) stub(rr)
time variable not set, use -tsset varname ...-
r(111);

// It seems I must set the time variable.

. tsset id year
       panel variable:  id, 1 to 2
        time variable:  year, 1980 to 1994

. rollreg y x1 x2 if id == 1, move(5) stub(rr)
. list id year rr_cons if id == 1 & year == 1985

     +----------------------+
     | id   year    rr_cons |
     |----------------------|
  6. |  1   1985   29.03253 |
     +----------------------+

// The constant is 29.03253 which differs from that for the ordinary OLS
// where _cons = 85.21505. What happens if I remove the if clause?

. drop rr*
. rollreg y x1 x2, move(5) stub(rr)
. list id year rr_cons if id == 1 & year == 1985

     +----------------------+
     | id   year    rr_cons |
     |----------------------|
  6. |  1   1985   29.03253 |
     +----------------------+

// Same problem.

I am not very familiar with the use of regression in a time series 
context. I am happy to accept that the results reported by rollreg are 
"superior" because they take the temporal nature of the data into 
account. However, is there a way I can force rollreg to generate the 
same results as those generated by a simple "regress" command?

Thanks for any help you can offer.

Cameron
*
*   For searches and help try:
*   http://www.stata.com/support/faqs/res/findit.html
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/