Chromatography Forum

I developed a quantitative GCMS method with 5 calibration points. The r2 of the calibration curve is sometimes <0.99. But is >0.99 when I eliminate one datapoint. The datapoint I omitted may be the highest calibration point or a point in between. I would like to know:
1. If there is any statistical calculation for choosing the outlier point to eliminate.
2. If there is any guideline stating whether omitting point in quantitative analysis is allowable and if so, how many point I can omit.

Please advise. Thanks.

Check out SW-846 method 8000C section 11.5 for one view.
http://www.epa.gov/epawaste/hazard/test ... 00c_v3.pdf

Note that removing points within your calibration range is not allowed. Reducing the range is OK. The total number of points required will vary with the model used, linear, quadratic, etc.

In regulatory analysis (general Pharmaceutical, HPLC) removing calibration points is a definite No-No.

It is permitted in bioanalytical methods but there are strict conditions and you still have to have 6 non-zero calib points after the removal.

Thanks Steve and JGK. The EPA guideline is rather stringent. I am analyzing human serum sample with GCMS with unweighted least square linear regression. Please enlighten me if there is any guideline for bioanalysis of clinical specimens or for toxicological analysis.
Many thanks.

If you cannot reliably generate a straight line fit using standards, why do you think that you can reliably generate accurate results for samples ? And if the method is not reliable, should you be using it in its present form, or troubleshooting it to make it reliable without having to prune the data ?

Peter

Personally, I absolutely agree with all the replies from those referring to regulated work. Even if your work is purely non-regulated research samples, you're skating on thin ice if you start removing calibration points to make things fit.

In the non-regulated environment in which I work, there are only two situations where I would delete calibration points:
(1) from the ends of a calibration curve which clearly extended too far (I've hit the limits of the detector in either direction; happens during method development).
(2) in the middle of a curve, where something has gone materially wrong, with evidence from more than just a bad data-point, i.e. the autosampler squashed a vial or pushed a septum into a vial, and that particular standard would have evaporated or injected incorrectly. In this case I will use the remaining points provided they aren't too widely spaced; even here I would report the problem. So far in my career I've only had to do this once, I think...

how about using quadratic fit?

ok to remove upper and lower points to reduce linear range but mid points not. our SOP is to make the std again which failed and re-run

Statistical tests for teh likelyhood of a data point being a true outlier (and, thus, properly ignored) are generally referred to ad Rho tests (17th Greek letterresembles: Ρ).

One example is Dixon's Q (don't ask why a Rho test is called a Q, I don't know).
As others have mentioned - compendial tests seldom, if ever allow for ignoring a result 9without assignable cause as to why it is off), even if you properly apply a Rho test and show that a result is probably an outlier.

Thanks Steve and JGK. The EPA guideline is rather stringent. I am analyzing human serum sample with GCMS with unweighted least square linear regression. Please enlighten me if there is any guideline for bioanalysis of clinical specimens or for toxicological analysis.
Many thanks.

http://www.fda.gov/downloads/Drugs/Guid ... 070107.pdf

See section VI

www.fda.gov/downloads/Drugs/.../Guidances/ucm070107.pdf

it is a guideline for bioanalysis from FDA, I am analyzing human serum sample too, and we follow this guideline for our bioanalysis.hope this helps