by
lmh » Fri Aug 28, 2015 1:50 pm
A couple of comments:
(1) You can't really have a real calibration point at zero concentration. You can force a one-point calibration curve through the origin, if you wish, but that's different (and that should give you a calibration curve that is really y = 500x + 0).
(2) If you have a curve like y = 500x+80 then the 80 must be the actual measured value in your zero-concentration sample. This means that you integrated a peak that wasn't there. It's possible the integrator parameters were set to such a sensitive level that they were integrating noise.
(3) This brings up the concepts of limit-of-detection and limit-of-quantification. Your results will not be reliable if the peak is not reliably bigger than noise. This is the limit of detection. They will also have large errors unless your peak is big enough that statistical random variations don't affect its area greatly. This is the limit of quantification. There are lots of defined ways to measure them, but that's not really the issue here. The issue is that the bottom end of your calibration curve, if it is from a zero-concentration standard, must, definitely, be below the limit of quantification, so you cannot trust measured values around this point. The problem is, how much higher does the concentration have to be, before you can trust it? Currently, you don't know, but you need to know.
(4) What you need to do is create a calibration curve with a series of points spread across the range of concentrations you intend to measure (for example, a dilution series from a high concentration).
(5) Another question is whether an error of 80 at zero concentration is relevant. If your calibration curve spans the region 100-1000 units on the x axis, and the curve is y = 500x + 80, then the actual measurements are 50,000 to 500,000, and 80 is merely an insignificant statistical error. If your calibration curve spans the region 1-2 units and you have this y-axis intercept, then the actual measurements presumably span the region 500-1000, and 80 is 16% of the low-end of the calibration range. That's not good.
(6) If you have a problem and 80 is significant, it may (as others have said) mean that your calibration is non-linear. This isn't necessarily tragic. Some detectors are linear, others do tend to non-linear curves, and every bit of software I've ever used can fit curves. If you need to fit a curve, you need enough points to be sure the curve is a good fit.
(7) But the bottom line is that whatever you do, your data are really only valid if they are somewhere between the lowest concentration standard and the highest. Anything outside this region is extrapolation, and will be increasingly inaccurate the further it strays from the standards. You are right to highlight the problem of negative concentrations for samples that don't reach the low end of the calibration curve, if the curve isn't forced through the origin. All this really means is that they are indistinguishable from zero. You can force a method to fail more elegantly by forcing the curve through the origin, but it doesn't make the data more reliable! Theoretically these low results should merely be reported as < LOD, but I do understand that in some fields, people expect numbers.