Re: Re: Re: regress versus fit - force throughzero/forceconstant term to zero

*To*: mathgroup at smc.vnet.net*Subject*: [mg61734] Re: [mg61613] Re: [mg61574] Re: [mg61493] regress versus fit - force throughzero/forceconstant term to zero*From*: ggroup at sarj.ca*Date*: Thu, 27 Oct 2005 05:54:28 -0400 (EDT)*References*: <acbec1a40510192358g44579892k5d78ab545b45e151@mail.gmail.com> <002901c5d556$c37f97f0$0401a8c0@achirana> <200510220911.FAA14997@smc.vnet.net> <200510230946.FAA10837@smc.vnet.net> <1154551111.20051027002059@aggarwal.ca>*Reply-to*: ggroup at sarj.ca*Sender*: owner-wri-mathgroup at wolfram.com

It was Thursday, October 27, 2005 at 12:20 AM, when I wrote: > On Sunday, October 23, 2005 at 05:46 GMT -0400, Chris Chiasson wrote: <snip> >> What is the best way to take into account the variance in the best fit >> parameters due to the fact that the model might be not 100% correct >> **and** the fact that each data point is probably not infinitely >> precise? <snip> Sorry, I didn't really answer the question. The regression is taking the into account all sources of variation in it's parameter error estimates. If you want to *separate* the contributions from various sources (ie how much of that variation is due to an inaccurate model vs noisy data), then your best bet is to model the noise and the expected form of the inaccuracy (for example DC offsets). You have to be careful that you don't add so many parameters that your model can fit any data set. A Bayesian analysis can help you quantify a model complexity parameter. If you have an estimate of the error bars (non-uniform) for your data points, you can also use the Weights option to modify the initial relative importance given to each data point while performing the regression.

**References**:**Re: regress versus fit - force throughzero/forceconstant term to zero***From:*Chris Chiasson <chris.chiasson@gmail.com>

**Re: Re: regress versus fit - force throughzero/forceconstant term to zero***From:*Chris Chiasson <chris.chiasson@gmail.com>