Re: st: nlsur quaids result depends on previous command

["Brian P. Poi" <brian@poiholdings.com>]

On 07/09/2011 02:09 AM, Johannes Kutsam wrote:
> Dear all,
> (and especially Brian Poi),
>
> I am estimating a QUAIDS demand system with nlsur quaids (POI 2008)
> (http://www.stata-journal.com/article.html?article=st0029_1 )
>
> I think found an error in either nlsur or quaids:
> The coefficients are dependent on the command called before.
>
> It can be reproduced using the example in sj8-4
>
> webuse food
> nlsur quaids @ w1 w2 w3 lnp1-lnp4 lnexp,ifgnls nequations(3) param(a1 a2 a3
> b1 b2 b3 g11 g12 g13 g22 g23 g33 l1 l2 l3) nolog
> nlsur quaids @ w1 w2 w3 lnp1-lnp4 lnexp,ifgnls nequations(1) param(a1 a2 a3
> b1 b2 b3 g11 g12 g13 g22 g23 g33 l1 l2 l3) nolog
> nlsur quaids @ w1 w2 w3 lnp1-lnp4 lnexp,ifgnls nequations(3) param(a1 a2 a3
> b1 b2 b3 g11 g12 g13 g22 g23 g33 l1 l2 l3) nolog
>
> Note that the 3 line with "nequations(1)" is of course wrong - because we
> have 3 equations. BUT once this line was executed
> The correct command  with "nequations(3) will yield wrong coefficients. (I
> am not saying one should run wrongly specified systems...but it happens when
> developing)
>
>   I suspect that some variables are not initialized correctly in quaids that
> could cause this problem.
>
> When I clear the dataset and load it again I again get the same coefficents
> as reported in sj8-4. So there is something temporarily saved by the call to
> nlsur quaids that confuses
> a subsequent call.

This is not a bug in -nlsur- or my QUAIDS code but an indication that if 
you write your own evaluator program and then call it incorrectly, 
-nlsur- could change your dataset.  As long as you don't subsequently 
(re-)save your dataset to disk, there is no permanent harm to the data.

In this case, the variables w2 and w3 are being overwritten by 
-nlsurquaids.ado- in the second -nlsur- call.  That's why your third 
call to -nlsur- doesn't work unless you reload the dataset.

Let's look at the second call to -nlsur- more carefully:

> nlsur quaids @ w1 w2 w3 lnp1-lnp4 lnexp,ifgnls nequations(1) param(a1 
> a2 a3 b1 b2 b3 g11 g12 g13 g22 g23 g33 l1 l2 l3) nolog

You've told -nlsur- that there is one equation and hence one dependent 
variable.  Therefore, -nlsur- is going to replace your first variable in 
the varlist with a temporary variable that your program will fill in 
with the fitted values of the dependent variable.  Thus, -nlsur- is 
going to pass along the following varlist to your program:

   __00**** w2 w3 lnp1 lnp2 lnp3 lnp4 lnexp

where __00**** is a temporary variable, the last four characters of 
whose actual name are generated at run time.

Now the version of -nlsurquaids.ado- that we are using assumes there are 
three dependent variables that it is to fill in.  So, -nlsurquaids.ado- 
is going to do -replace-s on the first three variables it is passed -- 
__00****, w2, and w3.  The ado-file was written assuming the first three 
variables are tempvars it is to fill in, but because we called it via 
-nlsur- with the wrong number of equations specified, it's going to 
overwrite w2 and w3.

When I'm working on something like this, I do everything inside a 
do-file, and by habit I always reload the dataset at the top of my 
do-files unless I have a particular reason not to.  That's why I've 
never come across this issue before.  But it is definitely an issue to 
keep in mind.

   -- Brian Poi
   -- brian@poiholdings.com


*
*   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/