Calibration curves

[cody841]

Hey I am working on a cannabis method for HPLC using Labsolutions.
I have created a calibration curve from 1 - 200 µg/mL and run it in triplicate. The results are linear, but then I find if I run the curve again as unknown samples the %recovery on the lower end is way out.

However, if I use exponential as the line of best fit, the %error drops for all points and the recovery is much more acceptable.

When modeling calibration curves and looking at the %error plots, how do you decide if there is a sample prep issue or systematic issue that is it ok to use different modelling? My error plots look like a very broad peak (high end has low error, then going towards middle increases and then drops to very low -% error at the low concentration end). I assume this means I am losing analyte as I go down, compounded with greater variation on smaller concentrations?

BTW I perform serial dilutions using micropipettes up to 300 µL each due to volume and cost of standards.

Thanks

[cody841]

Just to add some more info; for my lowest calibration point, the peak area is around 9000 and yet the curve built from that injection has a y-int of 19000. So you can see my problem at the low end. We can get negative values.

However this is completely corrected with exponential fitting.

I haven't even gotten to accuracy testing matrix issues yet...

[tom jupille]

What you're seeing is not unusual in a wide-range assay; ordinary least squares assumes that absolute errors are more or less constant across the entire range, whereas chromatographic methods typically have *relative* errors constant (so absolute errors are greater at the high end of the range). The problem can be sidestepped in a number of ways such as weighted least squares, transform to log-log, or polynomial fit. Check out Sections 8-10 of the Coleman & Vanetta "Statistics in Analytical Chemistry" series for more details. You can find links to the pdfs here:
- http://www.americanlaboratory.com/1403- ... lpi_4346=3