OLS - same sample, same regression - different results

[whoami]

Same sample, same regression - different results for coefficients beta1 and beta2 in Eviews.
why does this happen?

Should I check any background testing for this?
===================================================================

Estimation done on Nov 2nd:
---------------------------------------

Dependent Variable: Y
Method: Least Squares
Date: 11/02/09 Time: 16:27
Sample (adjusted): 2006M05 2009M09
Included observations: 41 after adjustments
Convergence achieved after 11 iterations


Variable Coefficient Std. Error t-Statistic Prob.


C 7.965833 3.779811 2.107469 0.0419
X1 0.004051 0.006373 0.635631 0.5289
X2 -0.041308 0.032648 -1.265238 0.2137
AR(1) 0.924773 0.048954 18.89055 0.0000


R-squared 0.936987 Mean dependent var 3.692195
Adjusted R-squared 0.931878 S.D. dependent var 0.920713
S.E. of regression 0.240308 Akaike info criterion 0.078674
Sum squared resid 2.136667 Schwarz criterion 0.245852
Log likelihood 2.387181 F-statistic 183.3943
Durbin-Watson stat 1.790986 Prob(F-statistic) 0.000000


Inverted AR Roots .92
==================================================================================================

Estimation done today:

Dependent Variable: Y
Method: Least Squares
Date: 02/05/10 Time: 16:41
Sample (adjusted): 2006M05 2009M09
Included observations: 41 after adjustments
Convergence achieved after 10 iterations

Variable Coefficient Std. Error t-Statistic Prob.

C 5.762153 3.245193 1.775596 0.0840
X1 0.010271 0.005786 1.775110 0.0841
X2 -0.023966 0.028353 -0.845288 0.4034
AR(1) 0.918397 0.047517 19.32792 0.0000

R-squared 0.941230 Mean dependent var 3.692195
Adjusted R-squared 0.936464 S.D. dependent var 0.920713
S.E. of regression 0.232077 Akaike info criterion 0.008976
Sum squared resid 1.992817 Schwarz criterion 0.176154
Log likelihood 3.815985 F-statistic 197.5228
Durbin-Watson stat 1.761103 Prob(F-statistic) 0.000000

Inverted AR Roots .92


============================================================================================

Thanks

[EViews Gareth]

Different starting values.

Starting values can (and often do) play a big role in the solution to non-linear estimation. You're estimating a non-linear specification, so starting values can make a difference to the estimates you get.

[startz]

Gareth's answer is usually the correct one in this situation. But you're getting very large differences in coefficients for a model that is piecewise linear in the structural coefficients and the AR(1) coefficient, and the latter has hardly changed. So, unless the X are very highly serially correlated I'm still surprised by your result. It's also odd that there's such a huge change in the SSR and log-likelihood.

On the other hand, least squares is unlikely to be broken either now or before....

You might try setting initial parameter values at the spot where your earlier run ended and see what happens.

[whoami]

Thank you for your replies. It is not a non-linear regression - I will check with the starting value settings.

[EViews Gareth]

You have an AR term - that makes it non-linear.