I would like to know the paramaters of that should be taken in account to claim the linearity of a analytical method or a instrument. What is the minimum value for R^2? What is the limit for the value of standard deviation and how to calculate this value? (Which variation is taken in account to calculate this?) Is it required that the intercept passes through zero?

rē>0.99 is a common requirement for defininf the linear range of either injector performance over a range of programmed volumes or for detector response over a range of concentrations. The line need not pass through zero (generally). To define a linear range, exceed it at both ends and pare away at the data until you're left with data whose regression results ina coefficient of determination >0.99.

I'm no statistician, but over the years I've learned that correlation coefficient is a lousy way to asses linearity, and it becomes totally meaningless unless the range of the data is specified. Try some calculations with "dummy" numbers, especially with a wide data range (like 25-150%) and you will see how far off some points can be and still give r2 > 0.99. I much prefer to asses the response factors at each level. This gives you a much better picture of the linearity of your system, in my opinion.

Is there a clear way to state on the linearity for given response factors?

I agree with the post by Anonymous about the value of correlation coefficient to assess linearity, especially in LC-MS where typical concentration ranges are very wide. According to FDA guidelines (see Guidance for Industry, Bioanalytical Method Validation, May 2001) calibration curves should be 1) within 20% of nominal concentration at LLOQ and 2) within 15% of nominal for all other standards. At least four out of six non-zero standards should meet the 15% criteria.

wHY ARE WE WORRIED ABOUT LINEARITY? wHAT IF THE DETECTOR RESPONSE IS NOT LINEAR? wHAT DO YOU DO THEN? sHOULDN'T WE USE MORE SOPHISTICATED TOOLS THAN A LINEARITY CHECK?

Sorry for this! The caps lock was on.

Agreed - rē is a blunt tool and requires careful examination of the data, but it is also generally understood (even if poorly) by a huge number of chromatographers. Keeping the individual points to within a percentage of the line is a good idea for an added criterion.

As far as nonlinear data goes, this is not generally a good idea, especially with a detector designed to exploit Beer's law, which deals with a linear relationship. Practically speaking, you could do some other sort of fit (2d order polynimial, exponential etc), but you would have to require a coefficient of determination much closer to 1 (0.9999999 or better) for it to have any meaning at all (from the perspective of a pharmaceutical analyst - things tend to be a little looser when examining plasma & other biological samples).

While we are on the subject of nonlinear curve fitting I was wondering if anyone had an opinion/comment about the use of quadratic equations to fit data from electrospray MS analyses. Our laboratory does a lot of PK analysis and our typical calibration range is from 1-5000 ng/mL. As you might expect we often see a saturation in response at the high end of the range (the curve flattens out) and we often exclude these points to observe adequate linearity. (We also use 1/x or 1/y weighting to better fit the low concentrations) However, I am aware of many people in the field who fit similar data with quadratic curves - the argument being that this type of equation is a better model for the nonlinear nature of analyte response inherent in the electrospray process. Any thoughts?(Apologies if this strays from the original thread)