I run quadchk after running a tobit to check the stability of coefficients.
Following
the manual, I use the rule that if the relative difference of coefficient
estimates is more than 1%
I reject the model as being unstable. However there are several things that
are happening that I would appreciate guidance on :
1.In some of my specifications, all my explanatory variables are 'stable' in
the way
defined above, but, sigma(u) and/or the constant turn out to be unstable.
Must I reject the specification as unstable because of this?
2. In one specification I get a missing standard error for sigma(u). What
could
this be happening (not in the quadchk, but in the main results)?
3. Can I infer anything at all from the 'amount of instability' of a
variable? i.e.
If a variable X is unstable by 2% or 20% in a specification, does
that tell me anything?
What if the specification nests another one where the model is stable,
does this convey anything? It obviously means
that the additional variable is somehow correlated with the unobserved
heterogeniety
of the panel (rho), but does it mean anything else?
Thanks very much!
-Jo
Jyotsna (Jo) Puri
Ph.D candidate, University of Maryland, College Park,
Department of Agricultural and Resource Economics.
