LOD determination

[Rick]

Hi!
I need your comments about this LOD determination (limit of detection).
I have read this type of determination in analytical methods:
LOD are calculated as 3x standard deviation from replicates of a blank sample fortified at a level wihich is 3 to 5 times the estimated LOD....for statistical reasons, if a calculated LOD using the standard deviation was less than 20% of the spike value, 20% of the spiked value was used as the LOD.

What do you think about that?

Thank for your comments

[aysal]

Hello,
In Chromatography, LOD/LOQ can be calculated by S/N ratio of injecting known amount of spiked samples. If there is no matrix effect, you may also use std injection just in solvent. S/N ratio of 3 is equal to LOD, 10 for LOQ meaning that you just start your calibration >LOQ value. For example, if you use 10 ppm for your calculation and you get 100 for S/N ratio, that means you have LOD value approximately 0.33 ppm and LOQ value 1 ppm. Next step is to check they are ok. Just inject 1 ppm and 0.33 ppm, and calculate S/N value. If they are consistent with each other (previous calculation), then you end the calculaton.
Theorytically, all calculations should result in almost same value. 0.33 or 0.35 or 0.30 ppm.

Ayhan

[gpronger]

LOD is a fairly dicey subject. A lot depends upon data usage. If you're doing work for a regulatory agency you need to determine how the agency wants this determined. The USEPA uses a procedure similar to what you originally described; typically 7 reps, spiked near what you expect to be your sensitivity limit, you then calculate the standard deviation and multiply by the student-t value (typically around 3).

Aysal's technique is common with the instrument manufacturers as well as academics. As I mentioned, depending upon data use either may be appropriate.

Since this is a forum around GCMS and HPLC-MS, a secondary issue is that assuming you're running in a full scan mode, is that as you go to lower concentrations, you will begin to have your secondary drop out. Both techniques have calculations based upon the ion current from (more than likely) your strongest ion. So, even trusting statistics, the point at which your primary ion is just statistically significant, your secondary ions are likely gone.

Once you start worrying about your secondary ions, you are somewhat on your own (I've not seen a protocol that specifically addresses this issue). What we have taken to doing is running the EPA study spiked at the lowest concentration we calibrate. This demonstrates precision and accuracy at that concentration. We then use that lowest concentration level as the lowest concentration we will report (adjusting of course for any sample concentration or dilution).

[Rick]

Thank you for reply.
Have you ever heard this.......
for statistical reasons, if a calculated LOD using the standard deviation was less than 20% of the spike value, 20% of the spiked value was used as the LOD.
Thank you!

[gpronger]

As Mark Twain stated; ""There are three kinds of lies: lies, damned lies and statistics."

Having said that, I have not heard of the protocol specifically, but what they are attempting to do is control the extrapolation of a LOD lower than reasonable (in this case, they are controlling it to 1/5th the spike level).

For example, when calibrating pesticides to low level (1 -5 PPB injected) it is not uncommon for the calibration to be linear, but with a significant X intercept (indicating a fixed amount of analyte "disappears" in the instrument or prep). Now, when I pick my spike level, I choose a concentration that will be around the lower end of my calibration (after sample prep). Because I have a pretty good prep department and autosampler, my injections have a high level of precision, and hence a very tight standard deviation. When I do the math, it is very easy for the calculated concentration to fall below the intercept. In other words, my calculated LOD could never occur due to the behavior of this particular analyte.

I've been in situations where after a MDL study did not achieve a low enough MDL, it was suggested to rerun the study, spike at a higher concentration to improve statistics, and thereby get a "better" LOD (though clearly the lab had data indicating that it could not see as low as they wished).

I hope this helps. As I said, I've not specifically heard of the protocol you've mentioned, but I do understand what they are attempting to avoid.

[lmh]

I can see why they're doing it this way, but I can also see a rather silly procedure coming out of it.

(1) Client specifies they need a LOD of at least 2 units.
(2) Laboratory therefore spikes with 3-5 times this value, say 5 times this value (we want to get a good signal!) = 10 units.
(3) Instrument detects it easily and the estimated LOD is 0.1 units.
(4) Laboratory therefore declare that LOD is 20% of spike, = 2 units, as requested by client.

In effect, given a good instrument, you are merely declaring that the LOD is as good as you expected, but you're testing it at 5 times the value you declare.

It would be better to test it at the declared value.

[Don_Hilton]

The problem with testgin the limit of detecation at the expected limit of detection is that at that level, one tends to spot noise and integrate it. And because the unaided eye is fairly consistant in picking out noise, the RSD's become surprisingly good at the limit of detection. Retention time reproducability becomes rather poor.

This at 5 to ten times the LOD, one has a real peak - not noise. Yes there is an assumption that the standard deviation does not change between the level of the LOD and the peak used to determine it. But I assure you that it is no worse and assumption that you can discriminate a peak close to the LOD to be sure that it is integrated and integrated correctly.

The EPA method does not go into much detail, but does have a caution about analytes with curvature in the low end of the curve.