Re: Quadratic regression
Posted: Tue Apr 08, 2014 6:54 pm
by dblux_
Is Empower perhaps swapping the variables? I believe one of our systems does it that way - incorrect statistically.
Let me answer instead of shaun78.
Empower and Excel gave the same equations.
The problem is only with Excel installed on the only one computer.
The reason of faulty behaviour is so far unknown.
I would carry out a thorougfull memory diagnostic test on that computer with something like Memtest.
Re: Quadratic regression
Posted: Tue Apr 08, 2014 7:46 pm
by Hollow
seems if your Excel don't like negative parameters...
I played with the Excel-Solver (add-In) and I obtain solutions (depending on the set accuracy) with only positiv numbers for the parameters a, b, c.
the best beeing:
80.496185 x2 +4134.034327 x + 94.088756
the residual sum-square of this solution is about 0.783% higher than of the 79.168... one.
With the parameter of your Excel, I got a sum-square which is about 111% of the 79.168 one.
"Nice" problem of the PC-World, even not satisfiable to work with.
Would be nice to hear some comment form Microsoft, how it would come?
Also would be a nice work for some math/stat geek to build a cube with the sum-square colored in relation to the global minimum, to show other local minima, which your Excel seems to be stucked to one.
Re: Quadratic regression
Posted: Wed Apr 09, 2014 2:29 pm
by dblux_
For comparison how ChemStation handles dataset in question:
Curve type: quadratic
Origin: ignored
Re: Quadratic regression
Posted: Mon Apr 28, 2014 12:20 am
by mrussell
There are data precision issues in Excel. As an example,I have a spreadsheet that does weighted regression by actually doing the calculations explicitly in excel. Its good to about 5 decimal places. But not exactly the same. I dont trust excel. The precision of a result can be affected by the way the calculation is performed. In your case though,it seems that its a error with your computer. Run prime 95 in stresstest mode.
HI all....
I've got a method where I need to use quadratic regression as the fit type for a calibration. A 5 point calibration curve is generated using the following points:
Code: Select all
Concentration (ug/mL) Peak Area
50 396287
71 717880
107 1363599
125 1757618
143 2247428
If I setup a quadratic fit in Empower, I generate the following equation for a best fit line:
79.2x^2 + 4400x - 11600
If I use Excel to generate an equation for the best fit line, the following is produced:
82.407x^2 + 3874.2x + 5919.9
I get the closeness of results on both the x^2 and x constant. I am confused at the difference in intercept as well as the total sign change. Clearly, results calculated using the two different approaches are not exactly the same; or even close for that matter.
For what it is worth, I can calculate the regression line on my TI-89, by hand, using Maple, and using Mathematica and I get the same equation that Empower generated.
My question is: why is there a difference? What is Empower et al doing differently that Excel is not?
TIA...