Y-intercept: linearity study in GC Solvent method validation

[sg]

Hi there,

I am currently encountering an issue for y-intercept during my GC method valiadtion work for determination of solvent DMF. As you know during method valiadtion for linearity study, the y-intercept against response at nominal concentration should be reported.

The R value I obtained is 0.9989. However the y-intercept value against response of nominal concentration is 22% which is too high to be acceptable.

Could anyone here to advise why such thing can happen? why good R value give such bad y-intercept? what does y-intercept against reponse of nominal concentration mean?

Thank you very very much!
sg

btw. the linearity range is from 60% to 140% of ICH limit for DMF. the LOQ is about 30% of ICH limit. I wonder if it is too close to LOQ level and the reponse is too low?

[krickos]

Hi Sg

Could you please share the method and which type of quntification you have used? (external standards or standard addition, how does sample/standard preparation look like).

You could have a sample matrix effect but hard to tell without the detailed information.

As for the concentration it is also hard to say without chromatogram/info about how much you put into the vial (30% of ICH limit 880ppm says "nothing" without knowing dilution/sample preparation).
However as a general comment, DMF is a high boiler (156°C) has a polar function and as such DMF is somewhat harder to analyse with headspace injection than a equivalent (response wise) non polar solvent with lower bp.

22% too high, I assume you mean a recovery of 122%?. What is the RSD at that level?

[chromatographer1]

Nominal concentration?

Your data is confusing. If you have an x-axis intercept which is negative, then your sample contains DMF and that moves your line according.

OR, your method contains an artifact which skews your line. This could be system DMF contamination which is not unusual if you have a cold spot or active sites present.

The easiest explanation is your dissolution solvent contains an impurity which coelutes with DMF in your chromatography system. If you are using DMAc as a solvent this is quite possible.

More information about your hardware and solvents=concentrations used are needed before a good trouble shooting evaluation can be offered.

best wishes,

Rodney George
consultant

[sg]

Hi,

The method is OVI method using external standardization. Agilent G1888A headspace sampler is used.

The system has no DMF contamination. DMSO is used as diluent. For the linearity study, I did not use sample solution but just use DMSO solution containes different conc. of DMF. Each conc. level I prepare three solutions. So the data used for linearity is from the mean of three results.

The regression is y=10.5343x-0.0722. The peak area for 100% of specification limit is 0.8111.

Since the DMF peak response is not high, I found that the integration can cause very big impact on the linearity especially at low conc. like 60% and 80%.

Here are the raw data:

Conc. (%w/w) peak area
0.0508 0.4832
0.0677 0.6151
0.0846 0.8111
0.1015 1.0098
0.1184 1.1760

The calculated LOQ is 0.03%w/w and 6 spiked sample at this conc gives %RSD 6.0 which is considered good.

Please let me know if more data needed.

Thank you!

[grzesiek]

from the data you've provided "the y-intercept value against response of nominal concentration is" about 9%. How did you calculate 22%?

"Each conc. level I prepare three solutions. So the data used for linearity is from the mean of three results." - how? dilutions of three main solutions?

"Here are the raw data: " - these are not raw data, this are calculated concentrations and mean areas, please provide raw data

[mbicking]

To answer some of your earlier questions:

Correlation coefficients (R-squared) reflect how closely the data lie along the least squares line. This parameter has nothing to do with the intercept. It only reflects how well the system responds to a change in the input value (concentration). But it is not a good single measure of calibration quality.

I am disappointed that regulatory and standards organizations still have not adopted a more statistically based aproach to evaluating calibration. A zero-intercept test is much more useful than these ratio calculations.

[sg]

Thanks guys!

Maybe someone can just explain to me what does it mean if the ratio of y-intercept to the response at nominal concentration is high. Does it mean the line is far from zero point? and why good R value gives high ratio result. Any relation between them?

Thanks!

[mbicking]

As I said, R-Squared is unrelated to the actual slope and intercept. It reflects how much of the change in response is due to the change in concentration (the rest being due to random errors). The fact that it might be better at the same time you have a certain intercept/slope ratio is just coincidence.

I have never used any formal comparison between the intercept and any other part of the curve. In an ideal world your intercept is very close to zero and you have a large slope (high sensitivity).

Perhaps at some point someone found an example where a certain relationship between the intercept and the nominal value had some meaning. But it is just as likely that someone just selected a value arbitrarily, and it somehow became an accepted practice, simply because it is easy to calculate and understand. (This is more common in the regulatory world than you might imagine.)

Assuming that your nominal value is at a relatively high response, a non-zero, and increasing, intercept means that your curve is changing, and could signal the presence of errors. But it really doesn't tell you much, and more investigation would normally be necessary. But nobody wants to do that because it involves statistics, and decisions, and it takes too much time.

OK, I'm in a cynical mood, so I will stop and let someone else pick up the discussion...

[krickos]

Hi

I took the liberty of feeding your data into excel and used an add-in for some linear regression calculations.

For the intercept I got this 95% confidence interval: -0.1814 to 0.0371.
So it does include zero but quite tilted to the negative side. More tilted (% wise) to one side than I am used to see.

What unit is the area in? mV*s? I am used to pA*s where a quite small peak is like 2,98745.

It could be that your method is too insensitive (area response wise) so that small normal variations have a larger impact, close to LOD/LOQ etc.

If you are in the pharma buisness, ICH Q2 R1 does actually recommend some linear regression analysis (deviation of the actual data points from the regression line etc) but not really any requirements.

[sg]

Thank you Krickos and mbicking!

Yes, the method is not that sensitive for DMF, quite close to LOQ level and the small changes affect quite significantly.

I think I have better understanding based on what you said...Thank you very much for your explanation!