Opened 14 years ago
Closed 9 years ago
#175 closed feature (fixed)
Display of covariance of parameters
| Reported by: | ajj | Owned by: | ajj |
|---|---|---|---|
| Priority: | major | Milestone: | Wish List |
| Component: | Analysis | Keywords: | |
| Cc: | Blocking: | ||
| Task: |
Description
Provide means of easily viewing covariance to warn of correlated parameters in fits.
Perhaps have an automated warning?
Change History (7)
comment:1 Changed 14 years ago by ajj
- Milestone set to Analysis Wish List
- Owner changed from srkline to ajj
comment:2 Changed 12 years ago by ajj
- Milestone changed from Analysis Wish List to Wish List
comment:3 Changed 11 years ago by srkline
comment:4 Changed 10 years ago by srkline
added a menu option in the SANS Models ->1D to show the correlation matrix. This brings up a panel that shows the values. along with the current parameter names. the help button links directly to the WM help file that explains the calculation. May want to add coloring for warning, if anyone uses it... Most important would be to make it more visible and in the users line of sight if there is a problem...
comment:5 Changed 10 years ago by srkline
Could the correlation matrix be added to the bottom of the wrapper panel, extending the whole panel, making it larger (taller) as needed? It could collapse/expand? It could be a lot of work for nothing?
comment:6 Changed 10 years ago by srkline
need to make a movie highlighting the covariance and the chi-squared map before I close out this ticket.
comment:7 Changed 9 years ago by srkline
- Resolution set to fixed
- Status changed from new to closed
Analysis items shelved
Function has been added: DisplayCovarianceMatrix?() in WrapperPanel?.ipf
No it's just how best to display this (and explain it to users). It could be part of a report (always, by default), always generated when the fit is done. It could also be generated manually.
Color the items in a table? What is the threshold? Repeat the boilerplate text from WM help file:
"A correlation matrix is a normalized form of the covariance matrix. Each element shows the
correlation between two fit coefficients as a number between -1 and 1. The correlation between
two coefficients is perfect if the corresponding element is 1, it is a perfect inverse
correlation if the element is -1, and there is no correlation if it is 0.
You should be suspicious of fits in which an element of the correlation matrix is very
close to 1 or -1. This may signal "identifiability" problems. That is, the fit doesn't
distinguish between two of the parameters very well, and so the fit isn't very well constrained.
Sometimes a fit can be rewritten with new parameters that are combinations of the old ones
to get around this problem."