Question about result summary

[peng gao]

Hi, all

 

I have a question about how to translate the final results.

 

After the fitting, we will get summaries like:

    " #Number of iterations : 5
     #Reduced chi squared  : 0.60083
     #Rw - value           : 0.153399

              1 #Correlations greater than 0.8 :

              2  #Corr(p[11], p[13]) = 1
                  #Corr(p[11], p[15]) = 1
                  #Corr(p[11], p[18]) = 1

                  #Corr(p[33], p[36]) = 1
                 #Corr(p[34], p[37]) = 0.856962
                 #Corr(p[35], p[38]) = -0.8422 "

 

Does "Correlations greater than 0.8"  mean that the model and the data are coincident?

 

"Corr(p[34], p[37]) = 0.856962" doesn't always appear. What does "p[34]" indicate? ( the parameter #34?)

 

Thanks.

 

Penmg Gao  

 

  

[chris farrow]

 

  

--
You received this message because you are subscribed to the Google Groups "diffpy-users" group.
To post to this group, send email to diffpy...@googlegroups.com.
To unsubscribe from this group, send email to diffpy-users...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/diffpy-users?hl=en.

Hi Peng,

You are correct that, "p[34]" refers to the parameter #34. 

Large parameter correlation means that multiple parameters have a similar effect on the fit. If the correlation is negative, then the parameters are anti-correlated (one goes down as the other goes up during a fit). For example, the scale factor and thermal factors are often correlated because they both affect the amplitude of the calculated PDF. The smaller the correlation, the better.

A correlation factor of 1 means that two parameters are not independent. For example, the data scale and phase scale are completely correlated in a single-phase fit. You should try to reduce parameter correlations by modifying your model, and certainly eliminate any parameters that are completely correlated (or anti-correlated). You may be over-parameterizing the lattice parameters or atomic positions.

Best,

Chris

--
Christopher L. Farrow
Associate Research Scientist
Applied Physics and Applied Mathematics Department
Columbia University
Office: 4265 Biomedical and Physical Sciences Building
       Michigan State University
       East Lansing, MI 48824-2320
Phone: (517)884-5677
Fax: (517)353-4500
email: clf...@columbia.edu
web: http://www.columbia.edu/~clf2121

[Thomas Proffen]

Adding to Chris' response. Essentially these are the values of the
correlation matrix coming from the least squares fit. Essentially a
large number close to one means the refinement parameters are no
independent and as Chris mentions, you will want to reduce the number
of parameters. The threshold 0.8 is somewhat arbitrary and comes from
the old PDFFIT (which I wrote ;-)). So what you want is no
correlations > 0.8 in your refinement.

Thomas