Linearity not going through zero

[aceto_81]

Hi,

our SOP states that we need to perform linearity 3 times in a validation.
Each time a calibration line was made.
Requirements for each calibration line:
R^2 > 0.99
%residuals between -5% and + 5%
and intercept equal to 0 with a t-test (0 should be in the confidence interval)

But now we have a very precise method, with some %residuals of 0.005%.
This means we have a very narrow confidence interval, which doesn't include 0.
If you calculate the actual deviation from 0, we get 0.05% which is much less that the precision of our method (eg if we perform 6x an assay, our deviations between them is more than 0.05%).

How do you perform linearity, or which specifications are you using?
Should 0 be in the interval, even such narrow?

Thanks

Ace

[DR]

Our LOD and LOQ determinations are usually far enough above zero that we don't worry about whether it passes through zero or not. For single level standard assays, it usually does not matter except that you someimes need to specify "linear through zero" in a processing method so that the computer will come up with a positive slope value. In these cases, the assay standard conc. is high enough and samples close enough to it that we can safely ignore any deviation in the trend line caused by sending it through the origin.

For detector testing, we look for an intercept within a certain range (±5% of calculated) but there's no column involved, just an acetone/water gradient.

[MSCHemist]

Intercept issues came up so manny times I spent a week studying the issue. here is what I came up with.

if your calibration range is large then errors at the top end can distort the low end. When you make a typical least squares line corrections to the Y value to get it on the line have the same weight at the top and bottom calibration points. It assumes homoscedacity in the errors (that you have the same abolute error at all calibration level rather than more often the errors are relative). So the equation is just as likely to make a correction of the response equal to .5ppm at .5ppm as at 500ppm. Forcing or including zero often helps to overcome this effect. A more valid way to to create a weighted calibration curve (Chemstation can do this) and have it inversely weighted by ammount.

I recommend reading Statistics and Chemometrics for Analytical Chemistry.

Here is another good link talking about curve weighing when things are getting screwy at the low end of the calibration curve

http://www.chromatographyonline.com/lcg ... ail/624978