accuracy

[gerapas]

i need your opinions about an issue regarding method validation and transfer. In my team,it's a rule that when we perform an accuracy study, we always find the line of best fit (Cadded vs Cfound), look if the R square is >0.999, if the line goes through the origin, the %y-intercept, if the slope is 0.98<b<1.02 etc. Do you do the same? I haven't read anywhere that i must do regression analysis when i perform accuracy studies. It depends on the concentrations i'm working on? I really wanna know what you are doing and WHY.

[tom jupille]

. . . we always find the line of best fit (Cadded vs Cfound), look if the R square is >0.999, if the line goes through the origin, the %y-intercept, if the slope is 0.98<b<1.02 etc. . . . I haven't read anywhere that i must do regression analysis when i perform accuracy studies.

It may be just a matter of terminology. Linear regression, usually via some form of least-squares model is how you find the slope, intercept, standard error, and coefficient of determination.

[lmh]

tom, completely off-topic, but very few people ever think about this: Least-squares best fitting is more all-pervasive than people tend to imagine. An average is actually a least-squares best fit of the values...

[tom jupille]

I think I see what you mean, but I would "split hairs" and disagree if you only have one set of data. Finding the "average" is, of course, the first part of what a least squares algorithm does.

[gerapas]

Let's make my question simpler. When you do an accuracy study,do you look only for the recoveries or you bother to find the line of best fit, R square , y-intercept etc?

[Blazer]

I only look for recoveries.

[HPLCaddict]

I know the methodology you're talking about as the "recovery function" - a linear regression of added concentration vs. measured concentration should give a line with slope 1 and intercept 0. I used this myself - in the past, not anymore.
I stopped to use it for several reasons:
- first and foremost it's not requested by any guideline I know
- no validation is like the other and your acceptance criteria should reflect the purpose of the method. If you look at some sort of assay in biological matrices, with complex sample preparations and other nice stuff, you might be really lucky getting recovery rates of 80% - imagine how the recovery function looks like in this case
- even with "simple" purity methods, if you look at the recovery at different concentrations you might see significant differences which have no real impact on the methods performance in real life
- I started to HATE certain validation characteristics like the ever-present correlation coefficient! Is a line with r2=0.9999 really "more linear" than one at r2=0.995? What's the rationale of your criterion r2>0.999? The same is true for 0.98<b<1.02 - why? I guess it's the same case in your company as it was in mine. Some time in the past someone implemented these criteria - and everyone kept on using them without really thinking about them .

Short answer: I only look at %recovery!

[gerapas]

the slope appears to be the overall recovery so it must be between 98-102%, but noone asks for it, am i correct?

[HPLCaddict]

No, the overall recovery does not have to be 98-102%! At least not in any case. It all depends on what that method is used for. There definitely are cases where 98-102% is a decent criterion (such as assay determination of a pharmaceutical product) but there are a lot of other cases where 98-102% does not make sense.
That's what I was aiming at: In my eyes it does not make sense to have a strict set of rules and acceptance criteria at hand that are used for every validation you're performing. Acceptance criteria should be set on a case-by-case basis.
Concerning the recovery function: As I already wrote, I'm not aware of any guideline requesting it.

[Alfred88]

Dear gents.
The first post is about two different things.
1/ Accuracy.
2/ Linearity.

When you do method validation, you need to check the accuracy of the method for the actual product, and cover the low, the normal, and the high concentrations.
The acceptance criteria may be different for the low concentrations, due to the wider spread at this level.
For example, error may be +/- 5% for the low concentration, and +/- 3% for the normal and high concentrations. These values must be practical, realistic, and based on the equipment and matrix.

Also, you need to check for the linearity of the method, to cover the same range as in accuracy, using at least five different concentrations. R (correlation coefficient) > 0.99 is okay. However, you may want to plot the points (x vs y) for visual check, and do an extra residual plot to confirm the "random distribution."

Alfred

[HPLCaddict]

If I understood the OP correctly, the first post is NOT about linearity, but only accuracy!
One method of interpreting your accuracy data is linear regression of the data and checking slope for deviation of 1 and intercept for deviation of 0. This has NOTHING to do with the the validation element "linearity" or a calibration line. Don't mix these up.

[gerapas]

i ve finally persuaded my colleagues not to do a regression analysis but check only the recoveries. By the way, do you establish LOD when you perform a study for degradation products(quantitative) or LOQ is sufficient?

[gtma]

I have seen validation data where the assay recoveries were within 99.0% to 101.0% from 80-120% level. However, the R failed...got 0.9990 with criterion of NLT 0.9995. What's the implication of a slightly lower R? In this case, I don't think it has a significant impact to the method accuracy. So, why set the R so tight when it doesn't reflect a significant bias in the method accuracy?