st: re: prais-winsten regression

[Kit Baum <baum@bc.edu>]

Evan said

I am regressing energy projection errors against GDP projection errors.
Both are measured as percentages and are pretty well bounded by +/- 0.2
(20%).  We choose one prediction length and analyze a time-series of
such projections (1983 projection of 1988, 1984 projection of 1989, 1985
projection of 1990, etc.)

We are running Prais-Winsten regressions to address autocorrelation.
However, every so often we get very large (and completely absurd)
coefficients on either the variable or the constant term.  These do not
arise when we use OLS, though we worry that those estimates would be
biased.  These huge coefficients arise despite the fact that both Does
anyone know of any reason this might occur in a Prais-Winston
regression?  Thanks.

Here is the output from one problematic regression:


tsset trend, yearly

prais pct_error gdp_nom_error, ssesearch

    time variable:  trend, . to 0103, but with gaps

Number of gaps in sample:  4
(note: computations for rho restarted at each gap)

Iteration 1:  rho = 0.8944 , criterion =  -.0259697
Iteration 2:  rho = 0.8944 , criterion =  -.0259697
Iteration 3:  rho = 0.8944 , criterion =  -.0259697
Iteration 4:  rho = 0.9282 , criterion =  -.02532865
Iteration 5:  rho = 0.9806 , criterion =  -.024677
Iteration 6:  rho = 0.9806 , criterion =  -.024677
Iteration 7:  rho = 1.0005 , criterion =  -.02401302
Iteration 8:  rho = 1.0005 , criterion =  -.02401302
Iteration 9:  rho = 1.0005 , criterion =  -.02401302
Iteration 10:  rho = 1.0005 , criterion =  -.02401302
Iteration 11:  rho = 1.0005 , criterion =  -.02401302
Iteration 12:  rho = 1.0005 , criterion =  -.02401302
Iteration 13:  rho = 1.0005 , criterion =  -.02401302
Iteration 14:  rho = 1.0005 , criterion =  -.02401302
Iteration 15:  rho = 1.0001 , criterion =  -.02400141
Iteration 16:  rho = 1.0001 , criterion =  -.02400141
Iteration 17:  rho = 1.0001 , criterion =  -.02400141
Iteration 18:  rho = 1.0000 , criterion =  -.02399868
Iteration 19:  rho = 1.0000 , criterion =  -.02399868
Iteration 20:  rho = 1.0000 , criterion =  -.02399868
Iteration 21:  rho = 1.0000 , criterion =  -.02399868
Iteration 22:  rho = 1.0000 , criterion =  -.02399868
Iteration 23:  rho = 1.0000 , criterion =  -.02399843
Iteration 24:  rho = 1.0000 , criterion =  -.02399843
Iteration 25:  rho = 1.0000 , criterion =  -.02399843
Iteration 26:  rho = 1.0000 , criterion =  -.02399843
Iteration 27:  rho = 1.0000 , criterion =  -.02399843
Iteration 28:  rho = 1.0000 , criterion =  -.02399841
Iteration 29:  rho = 1.0000 , criterion =  -.02399839


Can you say "unit root" ???

Why do you think that prais---which is basically an OLS technique  
coupled with an estimate of rho---is going to give a consistent  
estimate for an I(1) series? Use -dfgls- on the dependent and  
independent variables and you will likely find that you have a  
textbook example of what Granger & Engle call a spurious regression.



Kit Baum, Boston College Economics and DIW Berlin
http://ideas.repec.org/e/pba1.html
An Introduction to Modern Econometrics Using Stata:
http://www.stata-press.com/books/imeus.html


Begin forwarded message:


> I am regressing energy projection errors against GDP projection  
> errors.
> Both are measured as percentages and are pretty well bounded by +/-  
> 0.2
> (20%).  We choose one prediction length and analyze a time-series of
> such projections (1983 projection of 1988, 1984 projection of 1989,  
> 1985
> projection of 1990, etc.)
>
> We are running Prais-Winsten regressions to address autocorrelation.
> However, every so often we get very large (and completely absurd)
> coefficients on either the variable or the constant term.  These do  
> not
> arise when we use OLS, though we worry that those estimates would be
> biased.  These huge coefficients arise despite the fact that both Does
> anyone know of any reason this might occur in a Prais-Winston
> regression?  Thanks.
>
> Here is the output from one problematic regression:
>
>
> tsset trend, yearly
>
> prais pct_error gdp_nom_error, ssesearch
>
>     time variable:  trend, . to 0103, but with gaps
>
> Number of gaps in sample:  4
> (note: computations for rho restarted at each gap)
>
> Iteration 1:  rho = 0.8944 , criterion =  -.0259697
> Iteration 2:  rho = 0.8944 , criterion =  -.0259697
> Iteration 3:  rho = 0.8944 , criterion =  -.0259697
> Iteration 4:  rho = 0.9282 , criterion =  -.02532865
> Iteration 5:  rho = 0.9806 , criterion =  -.024677
> Iteration 6:  rho = 0.9806 , criterion =  -.024677
> Iteration 7:  rho = 1.0005 , criterion =  -.02401302
> Iteration 8:  rho = 1.0005 , criterion =  -.02401302
> Iteration 9:  rho = 1.0005 , criterion =  -.02401302
> Iteration 10:  rho = 1.0005 , criterion =  -.02401302
> Iteration 11:  rho = 1.0005 , criterion =  -.02401302
> Iteration 12:  rho = 1.0005 , criterion =  -.02401302
> Iteration 13:  rho = 1.0005 , criterion =  -.02401302
> Iteration 14:  rho = 1.0005 , criterion =  -.02401302
> Iteration 15:  rho = 1.0001 , criterion =  -.02400141
> Iteration 16:  rho = 1.0001 , criterion =  -.02400141
> Iteration 17:  rho = 1.0001 , criterion =  -.02400141
> Iteration 18:  rho = 1.0000 , criterion =  -.02399868
> Iteration 19:  rho = 1.0000 , criterion =  -.02399868
> Iteration 20:  rho = 1.0000 , criterion =  -.02399868
> Iteration 21:  rho = 1.0000 , criterion =  -.02399868
> Iteration 22:  rho = 1.0000 , criterion =  -.02399868
> Iteration 23:  rho = 1.0000 , criterion =  -.02399843
> Iteration 24:  rho = 1.0000 , criterion =  -.02399843
> Iteration 25:  rho = 1.0000 , criterion =  -.02399843
> Iteration 26:  rho = 1.0000 , criterion =  -.02399843
> Iteration 27:  rho = 1.0000 , criterion =  -.02399843
> Iteration 28:  rho = 1.0000 , criterion =  -.02399841
> Iteration 29:  rho = 1.0000 , criterion =  -.02399839
*
*   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/