WEF Discussion Forums
Laboratory Management and Technical Issues
Which of these two MDL's for Beryllium by ICPOES is "most correct"? We are following the EPA (40CFR136 Appendix 2?) procedure.
1. Using a concentration of 0.2 ppb gives an MDL = 0.04 ppb
2. Using a concentration of 0.1 ppb gives an MDL = 0.08 ppb
Both MDLs meet the criteria that Tested Concentration and MDL must be within a ratio of from 1:1 to 1:10.
Which MDL is "best"?
Appendix "2" is almost right, Keith! It's Appendix B.
You don't mention how you came up with the 0.2 and 0.1 ppb concentrations to use in the MDL study, but I guess one can assume it was by one of the following three procedures specified in Appendix B...
1. Make an estimate of the detection
limit using one of the following:
(a) The concentration value
that corresponds to an instrument signal/noise in the range of 2.5 to 5.
(b) The concentration equivalent of three times the standard
deviation of replicate instrumental measurements of the analyte in
(c) That region of the standard curve where
there is a significant change in sensitivity, i.e. , a break in
the slope of the standard curve.
In sub-paragraph 2 immediately following the quoted material above, App B says to prepare a solution that is 1 to 5 times, the estimated MDL. Without knowing what your estimated MDL was, I don't think anybody can tell which of your values...if either...is correct or "best."
Assuming that both your 0.1 and 0.2 ppb solutions meet the "estimated MDL" criteria, I would conclude that the 0.2/0.04 situation is "best". App B cautions that choosing an estimated MDL that is too low might lead to an inflated MDL, and one could use that as justification for not choosing the 0.1 ppb situation as being "best".
Keith's data actually are two points on a "continuum" that's highest at the smallest allowed spike concentration and should go to zero as the spike concentration gets larger and larger. The careful use of significant figures should make the disappearance occur fairly quickly. (Think about this before you climb all over my bones.)
I can't remember how the EPA arrived at its range of acceptable spike/MDL ratio, but it has to do with coming to some sort of happy medium. I'd be fun to do a whole bunch of studies and actually plot the calculated MDLs as the spike concentration varies. Maybe somebody already did this??
Added later: I forgot to address Keith's question. I agree with Perry.
To further my point, in looking through a binder of articles on detection limits, I found the following, taken from the September 2006 Application Notebook that comes as an insert with Spectroscopy magazine...
Following the procedure at 40 CFR 136 for nickel by ICP, the authors got the following two results:
IDL = 0.0025 ppm
MDL = 0.0018 ppm
One would expect that, as you add confounding factors such as reagents, digestion procedures, etc., precision should fall off. Not the case, here.
As Mark Twain observed, "there are lies, damned lies, and statistics."
We run into this with ICP/MS MDL determinations all the time. The analyses are so easy and fast to run, we run 10 standards for all metals and will usually see two to three different but acceptable MDL's for each metal. But as pointed out, at the MDL you have either zero or one sig fig (depending on if you are a chemist or statistician) so be careful of meanless sensitivity.
The problem with the MDL (and that has led to lawsuits and the recently closed Fed. Advise. Committee on Hg MDL) is the assumption that noise (as sigma, or %RSD) is constant over the range from zero to 5 times the MDL; it's not. Only the ASTM D19 method for "Detection Estimate" addresses this issue (that I know of) by modeling uncertainty with concentration at low concentrations allowing one to pick the "exact" point when the det. limit level is reached (usually 33% RSD for 3 sigma uncertainty).
As appendix B suggests, this mulitple MDL problem can be addressed by iteratively converging on the single correct point (output from one study is input to the next).
Unfortunately, there is a huge disconnect between the practical use of data and compliance with regulatory limits at levels pushing detection. Data even at 3 times the MDL has much more uncertainty involved that is generally recognized.