Re: st: Regression analysis with a minimum function on the RHS

["Sebastian van Baal" <s.vanbaal@arcor.de>]

Thank you for your suggestions! 
 
> On Fri, Mar 5, 2010 at 2:14 PM, Austin Nichols <austinnichols@gmail.com>
wrote:
> This seems likely to be problematic no matter what you do--typically
> the objective function should be differentiable in the parameters in
> these kinds of problems.  What is the theory that drives this
> specification?  Is there an alternative parameterization that is
> differentiable?

I admit it is a special hypothesis. My model is based on psychological and
microeconomic theory -- a mixture that creates all sorts of problems but is
also very interesting (to me). An alternative parameterization could be the
following: My original problem 

y = {b0} + min({b1}*x1 , {b2}*x2) 

is formally equal to 

y = {b0} + 0.5*[{b1}*x1 + {b2}*x2 - abs({b1}*x1 - {b2}*x2)]. 

Would you think that the second approach is better suited for estimation
with nl? 

> Steve:
> How about this approach?
> 1. run -sureg- to fit the regressions separately on x1 and x2.  Apply
> -constraint- first to get equal intercepts.
> 2. Use b0 + b1*X1  where   b1*x1 < b2*x2;  otherwise use b2*x2.

This seems to be a good approach I hadn't thought about. The results look
promising, but I will have to consult the literature before I decide on
that. 

Thanks again 
Sebastian

*
*   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/