nls and if statements

Hi All,

I have a situation where I want an 'if' variable to be  
parameterized. It's
entirely possible that the way I'm trying to do this is wrong,  
especially
given the error message I get that indicates I can't do this using  
an 'if'
statement.

Essentially, I have data where I think a relationship enters when a  
variable
(here Pwd) is below some value (z). I don't know that value, so I  
want to
fit it. The data is Pw, Tsoil, and Cfl.

nlstest=nls(if(Pw>z){Cfl~K*exp(-b*Pw)+c
	}else{
	Cfl~K*(exp(-a*Tsoil))*exp(-b*Pw)+c
	},start=c(K=5.5, a=0.1, b=0.1, c=2, z=5))

On May 17, 2012, at 10:06 AM, DWatts wrote:

Hi All,

I have a situation where I want an 'if' variable to be  
parameterized. It's
entirely possible that the way I'm trying to do this is wrong,  
especially
given the error message I get that indicates I can't do this using  
an 'if'
statement.

Essentially, I have data where I think a relationship enters when a  
variable
(here Pwd) is below some value (z). I don't know that value, so I  
want to
fit it. The data is Pw, Tsoil, and Cfl.

nlstest=nls(if(Pw>z){Cfl~K*exp(-b*Pw)+c
	}else{
	Cfl~K*(exp(-a*Tsoil))*exp(-b*Pw)+c
	},start=c(K=5.5, a=0.1, b=0.1, c=2, z=5))

Perhaps recasting as Boolean equivalent:

nlstest== <- nls({Cfl~ (Pw>z)*K*exp(-b*Pw)+c) +

                      !(Pw>z)*(K*(exp(-a*Tsoil))*exp(-b*Pw)+c)},
            start=c(K=5.5, a=0.1, b=0.1, c=2, z=5))

-- 

Still untested in absence of data.

David Winsemius wrote

Still untested in absence of data.

Thank you, and my apologies for not giving some sample numbers. Below is a
short subset of the data set I'm working with.

Cfl=c(2.61,4.21,2,2.75,7.47,1.2,3.24,12.49,2.37,3.28,4.3,2.61,2.75,2.92,3.78,2.25,2.84,3.33,2.39)
Tsoil=c(30.01,27.72,34.91,21.96,22.83,29.65,20.56,23.72,33.96,31.57,19.15,28.49,30.02,20.14,30.23,33.99,32.58,17.87,19.78)
Pw=c(30,29,24,22,22,41.2,-33.1,-49.21,29.6,24.5,-6.1544,42.8,63,55.5,43.3,38.5,87.2,-0.32,79.5)

Rui's code seems to elegantly do what I was attempting (and will try the
Boolian version next), but I'm having initial parameterization problems.
I've tried doing what I've done before, which is use the excel solver to
give me some simple early estimates of params, but it doesn't appear to be
helping in this case.

nlstest=nls(Cfl~K*ifelse(Pw>  z, 1, exp(-a*Tsoil))*exp(-b*Pw)+c,
	start=c(K=0.14, a=0.01, b=0.04, c=3.5, z=5), trace=T)

I'll keep chugging away at this. Thank you for your help!

If the above data is all you have, I wouldn't chugg too long.
Looking at plots of Cfl~Pw and Cfl~Tsoil says to me that you
have far too much variability to fit your model.

You might find this plot instructive:

   library(rgl)
   plot3d(Pw, Tsoil, Cfl, col=1+(Pw<5), size=5)


Peter Ehlers