# A question about capabilities of linest function

• From: Jim Dukelow <jsdukelow yahoo com>
• To: gnumeric-list gnome org
• Subject: A question about capabilities of linest function
• Date: Sun, 20 Feb 2011 23:58:11 -0800

I am using linest to attempt a polynomial fit in two numeric variables and two categorical variables (for which I have assigned +1 and -1 proxy values). The fit I am shooting for is up to fourth power in the numerical variables and 1st power in the categorical variables, with cross-terms. I have columns for the observed values of L, L^2, L^3, L4, F, F^2, F^3, F^4, L*F, (L^2)*F, etc., L*T, (L^2)*T, etc., F*R, etc. -- 42 terms in all. The observed values of my dependent variable are in the vector b8:b349. The observed values of my dependent variable have up to 10 significant figures. The observed values of the independent variables have one or two significant digits in the measured first powers (the original data) and up to 8 or 9 significant digits in the higher powers.
```
Values of L and F are normalized to live between -1 and +1.

I marked an array 5 rows by 42 columns, entered the command

linest(b8:b349,f8:au349,true,true),  and  held down  CTRL-SHIFT-ENTER.

```
More of less instantaneously, the array was filled with #NUM error messages and my command was now
```
{=linest(b8:b349,f8:au349,TRUE,TRUE)}(42,5)[0][0]

The Help pages are not much help in telling me what has gone wrong.

```
I recognize that using successive powers of L and F, rather than orthogonal polynomials, may lead to ill-conditioning of the calculation. Am I simply asking linest to do more than it is able, or is my problem elsewhere. Should I take this problem to R or Maxima or Octave?
```