[R] how lambda is computed in smoot.spline given _df_
wolski at molgen.mpg.de
Mon Nov 8 10:39:26 CET 2004
I posted some days ago a question concerning the computation of lambda in the smooth.spline function (which I repreat at the bottom of the mail) given _df_ .
Unfortunately the documentation is not clear to me. Maybee someone can help to answer in my view the basic question:
If the penalized log likelihood is L = (y - f)' W (y - f) + lambda c' Sigma c
how the _lambda_ in the above equation is computed if _df_ is given and _spar_ not?
And, Is there a way to define lambda directly?
I am usign the smooth.spline function.
I am not sure how the _df_ (degrees of freedom) parameter,
if set, influences _lambda_ in eq:
L = (y - f)' W (y - f) + lambda c' Sigma c
Is _df_ substituting tr(Sigma), if defined, in the equation: r = tr(X' W X) / tr(Sigma)
which is used to compute: lambda = r * 256^(3*spar - 1)?
And how _spar_ is set if not defined?
More information about the R-help