Re: Re: A step too far in regression fit ?

[Karl-Friedrich R. <mai...@gm...>]

On 20:59, pl...@pi... wrote:
> On 10/11/11 19:44, Hans-Bernhard Bröker wrote:
>> On 11.10.2011 14:04, pl...@pi... wrote:

>> A parameter value of zero is particularly bad in that case. No fit
>> parameter may be zero.
> _wrong_
> 
> The following works. I frequently set zero as a starting value if if
> makes sense.
> 
> baseºsedrift=0
> if (use_baseline) fit  baseline(x) datafile using 1:4 via
> base,basedrift
> 
>>
>> As it is, you pretty much explicitly asked for 'fit' to fail. No
>> wonder
>> it did.
> 
> If that were the case , I seem not to have come across that rather
> essential point it in the doc. Since you are wrong that may explain
> the absence of such information.

Zero is not exactly forbidden as a parameter, but the absolutes of
parameters mustn´t (or rather shouldn´t) be apart by a factor of
more than a few decades.

1e-30 is, well, more than thirty decades away from the values of
your other parameters, and zero even more so.

The fitting algorithm isn´t guaranteed to work well then, running
outside of it´s specifications so to say.

As HB said, if you don´t have an offset, you shoudn´t try to
determine it. ;-) Same goes for a negligible offset, because then
you won´t find it, but the noise in your data gets a variable fitted
to it. That can´t be good.

I´m not sure myself what the error message

Singular matrix in Invert_RtR

exactly means. It is clear that the correlation matrix cannot be
calculated with variables whose absolutes are too far apart, but
what is Invert_RtR ?

Regards,  Karl