Method Validation - Linearity

[Uwe Neue]

Hans: Non-random errors such as outliers can be determined with statistical tools as well. Of course, statistics is of no help if you messed up the entire experiment...

[tom jupille]

To paraphrase Taylor: there are three types of error:
- random error
- systematic error
- blunders

[HW Mueller]

Uwe, Tom,
succinctly put, with other words, "use your brain".

[bert]

Bert,
the data should be in J Chrom B, 678, 137(1996). Sorry, no time at present for more.

Hans, your paper is on my desk But, unfortunately you presented no data, just the result of the statistical (regression) analysis

So, no data to see what you mean...

Regards Bert

[syx]

When we use residual curve, how do we evaluate the linearity using the points? How far is too far?

[bert]

syx,

nice to see you join in again! As I wrote before, it depends on what samples you analyze and the guidelines you have/want to follow. According to the FDA ‘Guidance for Industry, Bioanalytical Method Validation’ (for biological samples as tissue and plasma), accuracy should be within ±15% of the nominal value, except at LOQ or below, where it should not deviate by more than ±20%. The EMEA demands are published in Volume VIII. I am working in the bioanalysis field, so if you do pharmaceutical analysis, maybe another member can help....?

Good luck,
Bert

[Alex Buske]

Syx,

On the residual plot it is not the point to get the distance of single points from mean. The distances are described by the r (or r^2, or some standard deviation).
Important is the shape: the points should scatter. If you see an arc, a sinus wave or some geometrical figure in general, than there is some deviation from linearity.

[HW Mueller]

bert,
sample data are in table 4, page 148, the explanation is in the text on p. 147 section (d). Again: The "eye" immediatly sees the upcurving (nonrandomness), Excel´s ANOVA did not.
Alex, did I understand you correctly? You look at the r^2 curve shape to elucidate nonrandomness, do you really have an example where something is evident after all these calcs which is not seen in the original data (but is significant in a practical not statistical sense)?
As mentioned before, I thought I saw something, somewhere, like a correlation figure which indicated nonrandomness.
Uwe, That stuff on outliers.... I remember that one can do a test to see whether somethinng is "officially" an outlier. If you don´t throw it out the statistics keep it in there, if you have enough good data it will also be obscured by correlation coeffs, etc.?

[syx]

I am so amazed how could my question become so complicated? Could I summarize the answer that the correlation coefficient is still the best parameter to measure linearity though actually it is not but a measure of correlation?

Bert, I've got severe headache...

[bert]

I've got severe headache...

As long as it is not significant....

[bert]

The "eye" immediatly sees the upcurving (nonrandomness), Excel´s ANOVA did not.

Hans, thanks for the tip. Nice data set, and indeed with no significant deviation of the slope from unity and intercept from zero. However, as mentioned before, according to the FDA ‘Guidance for Industry, Bioanalytical Method Validation’ (for biological samples as tissue and plasma), accuracy should be within ±15% of the nominal value, except at LOQ or below, where it should not deviate by more than ±20%. The lowest two points deviate 30 to 230%! This demonstrates the usefullness of the back-calculation and also, as you stated the misleading interpretation of r^2.

cheers Bert

[DR]

re: outlier tests - Judge Wolin specifically mentioned outlier tests (Dixon's Q, other rho tests) as being a "no-no" in FDA vs. Barr Labs ruling, so if you're in a Pharma industry Q lab, forget outlier testing. If you have the luxury of using outlier testing, be careful as there are numerous versions and it is just a matter of time before one finds one that delivers the desired answer...

To the original question, yes, I would generally recommend using r² but you do have to examine the plot to be certain that there is no curvature.

[JI2002]

A good source of information about calibration diagnostics can be found in American Laborary (Nov 2003, Feb 2004 and Mar 2004). These are three-part series on calibration daignostics by David Coleman and Lynn Vanatta.

There are seven basic steps in the procedure which include R^2 adj, residual plot, lack-of-fit test, p-value for the slope, etc. It's mentioned that R^2 adj is actually the weakest diadnostic tool. Because several of the seven steps require replica data for the concentrations (as Putnam does for his calibration), analytical chemists need to decide if it's worth the effort to run a calibration like that.

[bartjoosen]

can someone post the dataset?

[syx]

can someone post the dataset?

You mean: dataset for example?