The Origin Forum
File Exchange
Try Origin for Free
The Origin Forum
Home | Profile | Register | Active Topics | Members | Search | FAQ | Send File to Tech support
Username:
Password:
Save Password
Forgot your Password? | Admin Options

 All Forums
 Origin Forum
 Origin Forum
 95% Confidence limits out of fixed parameter range
 New Topic  Reply to Topic
 Printer Friendly
Author Previous Topic Topic Next Topic Lock Topic Edit Topic Delete Topic New Topic Reply to Topic

giordandue

Italy
Posts

Posted - 09/11/2024 :  03:43:22 AM  Show Profile  Edit Topic  Reply with Quote  View user's IP address  Delete Topic
Dears, I observe that once defined range of input parameters for user defined function fitting, the resulting 95% Confidence limits (LCL and UCL) for the same parameters are out of the defined intervals and without any sense... For esample, I fix the range of rock uniaxial strength 50<UCS<100MPa and similarly for other two independent input parameters, and I derive 95%LCL=-4080 and 95%UCL=4280 (unit=MPa).
Note that the limits of input intervals are consistent with the variability of input points to be fitted.
Thank you in advance for your help

YimingChen

1613 Posts

Posted - 09/11/2024 :  09:23:34 AM  Show Profile  Edit Reply  Reply with Quote  View user's IP address  Delete Reply
The calculation of the LCL and UCL for the fitting parameters is not related to the bounds set for the parameters. You can check the formula for calculating the confidence intervals (CIs) on this page.
https://www.originlab.com/doc/origin-help/nlfit-theory#Confidence_Intervals

James
Go to Top of Page

giordandue

Italy
Posts

Posted - 09/11/2024 :  09:46:47 AM  Show Profile  Edit Reply  Reply with Quote  View user's IP address  Delete Reply
Thank you James for your comment. In reality, it is still unclear to me how it is possible to get so different results for each dependent parameters than the relative input ranges. Just as an example, I enter data points to be fitted as some simple equation results y=f(x1,x2..) where x1, x2... are the independent values with a defined range, and of course I enter values in the range for each parameters.
In any case will try to understand better the formulation you suggested.
Go to Top of Page

giordandue

Italy
Posts

Posted - 09/11/2024 :  10:03:02 AM  Show Profile  Edit Reply  Reply with Quote  View user's IP address  Delete Reply
"..for each independent parameters"
Go to Top of Page

YimingChen

1613 Posts

Posted - 09/11/2024 :  10:44:08 AM  Show Profile  Edit Reply  Reply with Quote  View user's IP address  Delete Reply
I am getting confused. Assume we are fitting with function y = a*x1+b*x2+c. Are you fixing the range of the fitting parameters (a, b,c) or the independent variables (x1, x2)? When you say confident limits, are you talking about the CIs of the fitting parameters or the predicted y value?

Could you also share your project file here?

Thank you.

Go to Top of Page

giordandue

Italy
Posts

Posted - 09/13/2024 :  02:55:39 AM  Show Profile  Edit Reply  Reply with Quote  View user's IP address  Delete Reply
James, as you previously asked, I attach the file to be sure that there is not misunderstanding..
I input the ranges of some parameters (GSI, mi..) included in the function y=f(x) and I derive resulting LCL-UCL of these parameters extremely out of the ranges.

https://my.originlab.com/ftp/forum_and_kbase/Images/grs_test.opj
Go to Top of Page

YimingChen

1613 Posts

Posted - 09/13/2024 :  10:11:32 AM  Show Profile  Edit Reply  Reply with Quote  View user's IP address  Delete Reply
Large parameter error relative to the parameter value indicates that the fitting function is over-parameterized. Please refer to the page below for the explanation.

https://www.originlab.com/doc/en/Origin-Help/The_Reason_Why_Fail_to_Converge#Over-parameterized_functions

The bounds you set for the parameters only affect the fitted values of the parameters, but they do not impact the parameter errors.

James
Go to Top of Page

giordandue

Italy
Posts

Posted - 09/13/2024 :  11:04:58 AM  Show Profile  Edit Reply  Reply with Quote  View user's IP address  Delete Reply
Thank you again James...
Among other, it seems to me interesting that "over-parameterization does not necessarily mean that the parameters in the model have no physical meanings. It may suggest that there are infinite solutions and you should apply constraints to the fit process"..
Go to Top of Page

Sam Fang

293 Posts

Posted - 09/14/2024 :  07:57:07 AM  Show Profile  Edit Reply  Reply with Quote  View user's IP address  Delete Reply
Hi giordandue,

"apply constraints" means to fix parameters or apply constraints among parameters.

But your fitted result is not caused by over-parameterization, but your input data itself.

It is obvious that your input data includes three different types: rows 1-29, 30-58, 59-87.

When I fit these three segments separately, parameters are:
sigci mi GSI
1-29: 50 8 40
30-58: 75 10 50
59-87: 100 12 60

And all parameters' standard errors are very small, so LCL and UCL are very close to fitted parameters.

Therefore, we suggested you fit three segments separately instead of fit all together because three parts are quite different. Otherwise, residual sum of squares in the fit will be very large, and the standard error of parameters will also be very large.

Sam
OriginLab Technical Services
Go to Top of Page

giordandue

Italy
Posts

Posted - 09/15/2024 :  03:16:48 AM  Show Profile  Edit Reply  Reply with Quote  View user's IP address  Delete Reply
Hi Sam, thank you for your suggestion. I understand your point and I need to clarifiy the following. The analysis is just an exercise, with objective of estimating the 95% LCL for using in design the corresponding parameters (GSI, mi,sigci..). As you remarked, in place of experimental data to be fitted, I used the main equation y=f(x_gsi,mi,sigci) to derive the 3 sets of points for min-med-max parameters setting, and I launched the fitting analysis with reference to the resulting values. In the specific case, You suggest to fit separately the three segment (for min-med-max), but how consequently to derive the "overall" 95% LCL, if possible?
Go to Top of Page

Sam Fang

293 Posts

Posted - 09/18/2024 :  03:18:21 AM  Show Profile  Edit Reply  Reply with Quote  View user's IP address  Delete Reply
Hi giordandue,

If you can estimate the mean and standard error of y for each x according to your min-med-max y data, you can fit your mean data, with standard error as instrumental weight, and you can derive the 95% LCL.

If you know the distribution of your min-med-max y data at each x, you can also generate random samples, fit them separately, and perform statistics on fitted parameters to derive the 95% LCL similar to bootstrap method.

Sam
OriginLab Technical Services
Go to Top of Page

giordandue

Italy
Posts

Posted - 09/18/2024 :  05:00:34 AM  Show Profile  Edit Reply  Reply with Quote  View user's IP address  Delete Reply
Thank you Sam, I will try to follow your suggestions
Go to Top of Page
  Previous Topic Topic Next Topic Lock Topic Edit Topic Delete Topic New Topic Reply to Topic
 New Topic  Reply to Topic
 Printer Friendly
Jump To:
The Origin Forum © 2020 Originlab Corporation Go To Top Of Page
Snitz Forums 2000