[R] degrees of freedom question

[Greg Tarpinian]

    R2.3, WinXP


Dear all,

I am using the following functions:

 f1 = Phi1+(Phi2-Phi1)/(1+exp((log(Phi3)-log(x))/exp(log(Phi4)))
 f2 = Phi1+(Phi2-Phi1)/(1+exp((log(Phi3)-log(r)-log(x))/exp(log(Phi4)))

subject to the residual weighting

 Var(e[i]) = sigma^2 * abs( E(y) )^(2*Delta)


Here is my question, in steps:

 1.  Function f1 is separately fitted to two different datasets
     corresponding to two different dose response curves.  These
     fits are unweighted.
 2.  Function f2 is fitted to the pooled data such that the two
     dose response curves are assumed to differ _only_ in log(r).
     This fit is also unweighted.
 3.  The residuals from #2 are used to estimate an appropriate
     sigma^2 and Delta to use in weighting.
 4.  The functions described in #1 and #2 are refitted, but this
     time weighted using the information gathered in #3.
 5.  How many degrees of freedom should be allocated to the 
     weighted residual sums of squares?  (There are three such
     SSE's, one for each individual model, and one for the overall
     joint model)


Much thanks in advance,

   Greg