7 replies to this topic

[girdeux]

There is any form to do a lineal regresion that pass through origin (point 0,0) in the CP300? With out put many points (0,0).

[Lovecasio]

Sorry, but I don't understand your question. Could you explain more?

[Behnoud]

I have this problem as you!
I need Linear reg passing a certain (x,y) not only (0,0).

[girdeux]

Sorry, but I don't understand your question. Could you explain more?

Ok you imagin that we have some points (x,y) and we want to make a linear regresion with this points, you obtain y = 6,9 x + 9*10^-2 and you know that this function f(x) necessarily for x=0 -> y=0. This function muts pass trought point (0,0).

9*10^-2 it's small value but some times it's better obtain a regresion like: y=a*x; a=cte; with out b (origin of coordinates), I know that it's possible with exel, but with CP?

One solution is add much points (0,0) to the list of regresion, but I look for a more pretty solution.

[The_AFX_Master]

well, that 9e-9 is for sure a floating point small error.. who cares?.. remember that you need N+1 points to perform an n-grade regression.. if you have more points of that N+1, and one point have a small offset, for sure you'll get such small numbers on the B constant. i know that you can enforce a regression to pass over the origin in excel, but afaik.. isn't possible( imho useless because you put errors ) on the classpad

Girdeux, you know how least squares method works (minimos cuadrados )?, seems to be you not.... You can read it on a Numerical Methods book (As Chapra-Canale or Burden-Faires). Well, PAP is the pro of the numerical methods here, wait for him

[PAP]

Well, PAP is the pro of the numerical methods here, wait for him

Wow, I know something about numerical methods, but I never thought I'm a "pro"....

Ok you imagin that we have some points (x,y) and we want to make a linear regresion with this points, you obtain y = 6,9 x + 9*10^-2 and you know that this function f(x) necessarily for x=0 -> y=0. This function muts pass trought point (0,0).

First of all, I have to say that I don't use built-in linear regression much (in fact, not at all). However, linear regression uses the least squares method in order to obtain a "best" fitting curve f(x)=a*x+b. In other words, you don't have control over the fitting function, so you can't make it pass from 0,0 (b=0). Adding a data point (0,0) will not help in most cases, since the fitting curve don't necessarily pass from a data point.
There is a fast solution, but you need to use CPLua and LNA for this: use my LNA function LMFit, with f(x)=a*x as a model function, instead of f(x)=a*x+b. LMFit will return the correct value for the slope, a, and the fitting curve will pass from 0,0 (not near this point). If using CPLua and LNA is not a problem for you, try it (it's easy, just modify the example program XLMFit). If you have problems, send me your data, and I will post a small CPLua program that uses LNA to do what you want .

[The_AFX_Master]

@ PAP, at least on the forum, You know more about NM's than us.

[girdeux]

I don't havo much time but I have made a program with CPlua wich calculates the coeficients a,b for y=a+b*x, the R^2 and the coeficient b for y=b*x.

It's all ok, but I get a bad R^2, for lines that cross the origin. I use this ecuations, can someone tell me if they are bad or if I should use other ecuations?

n = number of points
mx=mean(xi)
my=mean(yi)
mxy=mean(xi*yi)
mx2=mean(xi^2)

sy2=sum( (yi-my)^2 )/n
sxy2=sum( (xi-mx)*(yi-my) )/n
b=mxy/mx2
R^2=b*sxy2/sy2