| Author |
 Topic  |
|
|
jospector
8 Posts |
 Posted - 12/10/2008 : 3:06:39 PM
|
Origin Ver. 8 and SR 1
Operating System: XP
I have a bunch of data sets that a histograms. Some of them are clearly two Gaussian peaks, and others are clearly one. My question arises on my data sets that may be EITHER 1 or 2. I can't seem to find a way to discern the number of peaks. Is there a correct statistical test for this, and if so how would I implement it in Origin ? Any help is greatly appreciated..
thanks,
-js- |
|
|
VincentLiu
China
Posts |
|
|
easwar
USA
1965 Posts |
Posted - 12/12/2008 : 12:27:38 PM
|
quote:
Is there a correct statistical test for this, and if so how would I implement it in Origin ? Any help is greatly appreciated..
Hi,
First of all, it would help if you 'know' where the second peak could/should be. Then you can ask Origin to fit with two peaks, and in the process fix the peak center parameter, or at least put lower and upper bounds on peak center based on where it is expected to fall. This can help with obtaining some physically meaningful peak location.
Now, your primary goal appears to be to determine if one-peak-fit or two-peak-fit are statstically better. You cannot do this, at least now, with Peak Analyzer.
You can instead use NLFit tool, as follows:
1. First fit your data with one gaussian peak and create report
2. Then bring up NLFit tool again on your data, select Advanced from the list on the left under Settings tab
3. Set Number of Replicas to 1 (so a total of two peaks)
4. There are different peak finding methods such as 2nd derivative available under drop-down in that page, so those could be used to try find the 2nd peak location. But if those fail, or where they find the 2nd peak is not acceptable, click the Parameter tab and put in "expected" values for 2nd peak. Also put in bounds/constraints etc as explained above
5. Finish the 2nd fit to create report, make sure for both fits, the fit converged (does not make sense to compare if did not converge)
6. Then use Analysis->Fitting->Compare Models menu item. In the dialog that comes up, select the two fit sheets.
This Compare Models tool will then look at the fit statistics and use an F-test and an AIC test to output a statement (footnote in this tool's report) as to which model is statistically the better choice. Note that it is not appropriate to just compare R^2 etc as Degrees of freedom etc are involved - model with more parameters typically give lower R^2 etc as there is more "wiggle room", but that does not mean it is a statistically better model.
Hope this helps.
Easwar |
 |
|
| |
Topic |
|
|
|
number of peaks ?