[6412] in s-news-athena
Re: nls errors?
daemon@ATHENA.MIT.EDU (Douglas Bates)
Wed Feb 8 11:34:55 1995
Date: Wed, 8 Feb 95 10:13:36 -0600
From: bates@stat.wisc.edu (Douglas Bates)
To: cyn@odin.mda.uth.tmc.edu
Cc: s-news@utstat.toronto.edu
In-Reply-To: <9502081603.AA18251@odin.mda.uth.tmc.edu> (cyn@odin.mda.uth.tmc.edu)
>>>>> "CS" == Cyndi Smith <cyn@odin.mda.uth.tmc.edu> writes:
CS> Does anyone out there know how the nls function calculates "Standard
CS> Errors" and "Correlations"? The errors it gives us seem kind of
CS> low... Thanks.
The calculations of standard errors and correlations are based on a
linear approximation to the model evaluated at the parameter
estimates. You can try to assess whether the standard errors are
reasonable by profiling the model using profile(fitted.object).
Profiling involves reoptimizing the sum of squares conditional on
parameter components holding specified values. When successful, the
profile function returns a list of data frames with columns labelled
"tau" and the parameter names. One can construct a
profile-likelihood-based confidence interval on a parameter by
relating the tau values to a t distribution.
The linear approximation technique is a "standard" technique. The
profiling approach is described in chapter 6 of "Nonlinear Regression
Analysis" by Bates and Watts (Wiley, 1988).
