Explanation of regression weighting

[Ruben]

Hi All,

I am working with a Varian Saturn 2000, and am analysisng for environmental pollutants. Although I have done some reading on regression weightings I amstill confused. Is there an easier explanation to understand when to use the different weightings specified in the software?

Thanks

[Don_Hilton]

If you your run replicates of each level of a calibration curve and then plot RSD for an analyte as a function of concentration, you will see a curve with high RSD values at the end with low concentrations and low RSD values at higher concentrations. The form will be approximately y = 1/x. If the major source of variability in the results is background noise, it will be constant and the signal will be proportional to the sample concentration. Signal too noise will then be proportional to 1/concentration. -- Which is what you see in the plot, as RSD is a measure of relative noise.

When you weight your regression with a 1/concentration weighting, the resulting curve takes into account the fact that there is more noise at the low end of the curve - and should give you a more robust curve.

For the difference between using 1/x and 1/x^2, I would suggest the recent series on calibrations that has been running in LC-GC magazine (I believe this is the one - if I discover I'm thinking of the wrong magazine when I get to work on Monday, I correct this.) and can be viewed on line. And it goes into a longer (and better) discussion than I have given here - so even if 1/x vs. 1/x^2 is not the question, it is still worth digging up.