Estimate does not handle unknowns as it used to do

[jimarin]

Using tolbase/tol version 2006-09-05 (and also reproduced for version 2006-09-22)
we have experienced the following:

• having unknowns in the data produce errors when processing estimate function.

• the unknowns are not corrected and they appear on all estimation series (residuals,etc...)

And

• this does not happen when using older versions like 2006-08-17

Item more: the behaviour of tolbase.exe seems to be different from tol.exe, some times the error appears for tol.exe and does not appear in tolbase.exe.

An estimate isolation file has already been given to josp for his study on this, and the same file could be given on demand to any developer that should look onto this problem.

The problem is labelled as major cause our application typically deals with series with unknowns.

Perhaps it is interesting to note here that this error seems to be going on since bugs 424 and 428 were solved.

Thanks.

jima

[Víctor de Buen Remiro]

We are working about this bug
Thanks to report it

[Víctor de Buen Remiro]

Problem has been fixed in CVS.

This is an old problem that remained hidden due to the invalid treatment of the real values represented in floating comma like #IND (indetermination) and #NAN (Not A Number) which previously they did not treat like omitted. When it was fixed this appears problems like this that before were not seen.

In this case the problem is that the values of the interruptions could not be calculated, probably due to an error of numerical type in the partial matrix of the PI-weights.

To solve it, it is enough with controlling that the resulting matrix does not contain omitted so that the estimator acts correctly.

/
LO MISMO EN CASTELLANO
/
El problema ha sido corregido en el CVS.

Se trata de un viejo problema que permanecía oculto debido al tratamiento indebido de los valores reales representados en coma flotante como #IND (indeterminación) y #NAN (Not A Number) que anteriormente no se trataban como omitidos. Al arreglarse esto aparecen problemas como este que antes no se veían.

En este caso el problema es que no se pudieron calcular los valores de las interrupciones debido probablemente a un error de tipo numérico en la matriz parcial de los PI-weights.

Basta con controlar que la matriz resultante no contenga omitidos para que el estimador actúe correctamente.