Putting Limits on Coefficients

[cmgconsulting]

After reading your discussion on Putting Limits on Coefficients thread, I am following the code you posted below:

Code: Select all

equation eq2.ls y = c(1) + (2*@logit(c(2))-1)*x

My equation is:

log(ssu) = c(1) + (-3.5*@logit(c(2))-3)*log(price) + c(3)*log(acv) + c(4)*(items) + c(5)*promoacv + c(6)*promounits +c(7)*trend + c(8)*easter

however, it returns a singular error. If I replace the (-3.5*@logit(c(2))-3) with c(2), the equation works just fine. I've tried using exp(c(2)) just to see what happens, and that causes problems as well. Any help?

[EViews Gareth]

Is it a singular matrix error, or a Warning: Singular Covariance message?

[cmgconsulting]

"Near Singular Matrix"

[EViews Gareth]

Just means that EViews can't find a solution to the problem.

[s_pushparaj]

Is this due to the fact that the non-linear estimation can't be done using least square procedure. However there are other non linear estimation procedure available for this sort of estimation.
- Pushparaj

[EViews Glenn]

(Re: the previous posting by s_pusharaj) Lest anyone get the wrong impression, I'd just like to point out that the EViews least squares procedure certainly does perform nonlinear estimation.

If I had to guess at what is going on here I'd suspect that it's a starting value issue. We will use whatever is in the coefficient vector to initialize the nonlinear estimation. When you switch from estimating with the coefficient in levels to a transformation there may be problems with getting derivatives which cause singularities. The most obvious hint that this is the case would be if the "Near Singular Matrix" error comes up immediately when you estimate the equation (as opposed to running for a number of iterations first).

If this is the case, playing around with the starting values may help. Translating from the unrestricted coefficient using the inverse of the transformation would be an obvious first approach.