-
- Posts: 1680
- Joined: Sat Aug 23, 2008 12:04 am
There is an assumption here - that a calibration curve is linear from zero up to your lowest standard. This is a separate issue from weighting the values in the regression calculation.
One thing that happens, particularly in trace analysis is that a small portion of an analyte can be lost to active sites in the chromatographic system. And, careful examination of the low end of a calibration curve may reveal that the curve hooks and passes through zero. If you do not run standards at this low level, you may see a line that misses the origin with no signal at some real, but minuscule concentration.
There are other reasons for bias - and these may result from other physical processes in handling the sample or the detection method used by the integrator.
This is why we do things like validation studies and determine linearity, bias, and method precision. The studies are a pain. And if you want to give a good answer to the boss, pulling out the validation study folder is the best answer on why you set up the method the way you did.
And while we are running samples across our range of interest, a couple of other things. First if you run multiple injections at each calibration level, you will notice that the RSD for each point is proportional to the sample concentration across most of the range - thus the reason for the use of inverse concentration weighting for a fit. At the limit of the curve, noise introduced by the instrument becomes larger than noise associated with the analyte and this relationship falls apart.
And, at that low end of the curve where noise begins to dominate, one can use a student's-t test to show that one is not able to discriminate a value from zero down at that end. This is where the method is not good for making measurements.
And the last thing to note - and is seen by folks who do trace analysis of difficult compounds like pesticides: There is what is called the matrix enhancement effect. Those active sites I mentioned above... If you have a matrix that can tie up the active sites (like real life sample extracts) all of the pesticide or other analyte will survive to the detector, unlike standard solutions, which have no matrix (other than solvent) in them. Thus, for some analyses, we make our calibration curves in matrix - because of that troublesome intercept. And this is why we run spiked samples in our validation studies -- to demonstrate that matrix enhancement is or is not an issue.
One thing that happens, particularly in trace analysis is that a small portion of an analyte can be lost to active sites in the chromatographic system. And, careful examination of the low end of a calibration curve may reveal that the curve hooks and passes through zero. If you do not run standards at this low level, you may see a line that misses the origin with no signal at some real, but minuscule concentration.
There are other reasons for bias - and these may result from other physical processes in handling the sample or the detection method used by the integrator.
This is why we do things like validation studies and determine linearity, bias, and method precision. The studies are a pain. And if you want to give a good answer to the boss, pulling out the validation study folder is the best answer on why you set up the method the way you did.
And while we are running samples across our range of interest, a couple of other things. First if you run multiple injections at each calibration level, you will notice that the RSD for each point is proportional to the sample concentration across most of the range - thus the reason for the use of inverse concentration weighting for a fit. At the limit of the curve, noise introduced by the instrument becomes larger than noise associated with the analyte and this relationship falls apart.
And, at that low end of the curve where noise begins to dominate, one can use a student's-t test to show that one is not able to discriminate a value from zero down at that end. This is where the method is not good for making measurements.
And the last thing to note - and is seen by folks who do trace analysis of difficult compounds like pesticides: There is what is called the matrix enhancement effect. Those active sites I mentioned above... If you have a matrix that can tie up the active sites (like real life sample extracts) all of the pesticide or other analyte will survive to the detector, unlike standard solutions, which have no matrix (other than solvent) in them. Thus, for some analyses, we make our calibration curves in matrix - because of that troublesome intercept. And this is why we run spiked samples in our validation studies -- to demonstrate that matrix enhancement is or is not an issue.
