Help on constrained regression

1.  Have you made normal probability plots of the data 
(function "qqnorm")?  This could reveal outliers or a need for a 
transformation. 


           2.  Have you considered using "lme" with "corAR1", allowing 
"a" to be negative?  The examples in the help file for "corAR1" show how 
to estimate this model with a different "b" for each series but with one 
"a" shared between all series.  Then an analysis of residuals should 
help you identify which series seem least consistent with the model. 


           3.  Have you considered dropping one observation in each 
series and using "lme" to estimate a distinct "a" and "b" for each 
series, allowing  "a" to be negative?  If the grand mean of the 
estimates for "a" are positive, this will reduce the number of negative 
"a" values by shrinking them all towards the grand mean.  Then you can 
look more carefully at the series that still have negative "a" values.  
They will either contain outliers, or you will likely identify some 
other series to include in the model so it becomes something more like 
the following: 


           y[t] = a*y[t-1] + b + c*x[t] + e


      As your comment indicated, the seriously negative estimates for 
"a" suggest some substantive violations of the assumptions you used to 
conclude that "a" should be positive. 


      Hope this helps. 
      Spencer

I did. The problem was the underlying process that's negative AR(1).
So I just have to find other way to model it. Thank you.

On Thu, Jul 2, 2009 at 7:31 PM, spencerg<spencer.graves at prodsyse.com> wrote:
  

    Have you considered writing the model in terms of log(a) = g, say:

         y[t] = exp(g)*y[t-1]+b+epsilon?

    With this, you could estimate "g" and "b" using "nls".  With multiple
series, you could use the "nlme" function in the "nmle" package.  For the
"nlme" package, an excellent reference in Pinheiro and Bates (2000)
Mixed-Effects Models in S and S-PLUS (Springer).

    Hope this helps.
    Spencer Graves

R_help Help wrote:
    

Hi,

I have an AR(1) model

y[t] = ay[t-1]+b+epsilon

I'm trying to force a to be positive. So I did the constrained
regression with constraints 0 < a < 1. I used pcls in package mgcv.
However, I found that the solution is not so stable. Most of my lag 1
autocorrelation is negative. Forcing a to positive value makes the
optimizer to stick a to the boundary value. All it does is varying b.
I there anyway to solve this problem? I think the problem might be due
to my initial value is not a smart choice.

Thank you.

adschai

_______________________________________________
R-SIG-Finance at stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-finance
-- Subscriber-posting only.
-- If you want to post, subscribe first.