Frequently asked questions

Questions and answers about precision, trueness and accuracy topics.

 

Question: Why does the replicate measurement in case of precision assume that the replicates are “appropriately independent” but in case of linearity replicates can also mean that samples are run for example in triplicate (from the same vial)?

Answer: The way you have to plan your replicates depends on the purpose that you have. Carrying out sample preparation once and thereafter injecting multiple times is a valid activity, but it does not yield the precision of the method but only the precision of the LC-MS analysis part sr_LCMS. And importantly: sr_LCMS is smaller (often significantly smaller) than sr.

In order to obtain the precision estimate of the method the replicate measurements have to be carried out in such a way that all method's steps are included.

When determining linearity – carrying out triplicate injections of calibration solutions serves the purpose of reducing random scatter of data points caused by the LC-MS system on the graph. This approach does not enable reducing scatter caused by preparation of the calibration solutions. However, the scatter caused by the LC-MS is usually significantly larger than scatter caused by preparation of solutions.

Question: How to calculate representative bias when you have many bias values found in the validation eg for different fruits and vegetables or concentration levels?

Answer: The answer to this question to some extent depends on how the obtained bias data will be used. Here are some situations together with general recommendations. --- If bias is needed – either for reporting, correcting or for taking into account as uncertainty component – for a specific fruit or vegetable, use the bias for the specific fruit or vegetable. --- If bias estimate is needed for demonstrating that method's bias is sufficiently low for all matrices, then it is suggested to present the worst bias, if the worst bias is still tolerable. If it is not, then a closer look should be taken at the data. --- If bias estimate is needed for evaluating general measurement uncertainty of the analysis method (using the  Nordtest approach) then it is suggested to use the root-mean-square bias. A rather good general guidance material is a recent article about bias in journal Bioanalysis (ref 22 in the course). Although its title specifically refers to clinical chemistry, it is actually very relevant for any chemical analysis.

Question: Section 6.3. says “In the case of correcting results with bias, the uncertainty of bias has to be included into the result’s uncertainty budget. It can happen that the uncertainty of correction is not much smaller than the uncertainty due to possible bias.” However, one criterion has to be fulfilled in applying correction. "Useful reduction of combined uncertainty is achieved (as compared to inclusion of the possible bias into uncertainty)." It seems that even though correcting the bias for the final result, the uncertainty of the bias (before applying correction) still has to be taken into uncertainty budget. If so, the final combined uncertainty should not be reduced, and should be the same no matter the correction is applied or not. Could you please help clarify?

Answer: If there is (possible) bias in the result then there can be two situations, depending on, whether the result is corrected for bias or not.
--1-- If one corrects the result with bias then the uncertainty of the correction has to be included in the uncertainty of the result. That uncertainty can sometimes be large and if we speak about possible bias (i.e. if it is not sure whether there actually is any bias at all) then the uncertainty of the correction can be larger than the correction itself.
--2-- If one does not correct for bias then one should include the uncertainty due to possible bias into the uncertainty of the result. In both cases the (possible) existence of bias introduces additional uncertainty, whether we correct or not. The point now is that the uncertainty increase when correcting for bias should be smaller than in the case when the result is not corrected. Let us look at two examples.
--- Example 1 --- Suppose we have determined the content of a contaminant in a sample C = 120 μg/kg. Let us assume that the joint uncertainty contribution of all other (i.e. except bias) uncertainty sources is 10 μg/kg. Let us assume that we have determined bias and found that it is -20 μg/kg with standard uncertainty 6 μg/kg. In this case, if one would correct for bias one would end up with the result 140 μg/kg and combined standard uncertainty 12 μg/kg. If bias correction would not be carried out then the result would be 120 μg/kg and combined standard uncertainty would be 23 μg/kg (the uncorrected bias as well as its standard uncertainty have been included as standard uncertainty contributions). Obviously in this case bias correction is justified – uncertainty decreases by roughly two times as a result of correction.
--- Example 2 --- Suppose we have determined the content of a contaminant in a sample C = 120 μg/kg. Let us assume that the joint uncertainty contribution of all other (i.e. except bias) uncertainty sources is 20 μg/kg. Let us assume that we have determined bias and found that it is -10 μg/kg with standard uncertainty 9 μg/kg (this means that it is actually a possible bias – we do not know for sure if there is any bias at all). In this case, if one would correct for bias one would end up with the result 130 μg/kg and combined standard uncertainty 22 μg/kg. If bias correction would not be carried out then the result would be 120 μg/kg and combined standard uncertainty would be 24 μg/kg. In this case bias correction is not justified, because the uncertainty remains almost the same and we do not know whether there actually is any bias at all.

Question: Is it possible to use one sample t test instead of zeta score to evaluate whether two results are in agreement?

Answer: It is more correct to use the zeta score, because of two reasons:
--1-- The zeta score takes into account the combined standard uncertainty instead of just scatter of results. This way it accounts for both random and systematic effects.
--2-- The zeta score takes into account also the uncertainty of the reference value. As a result of these things it is possible (and not rare) that the reference value is not within the range of the individual measurement values (especially if the individual values were obtained under repeatability conditions) but at the same time the average measured value and reference value can still be considered to be in agreement.

Question: What is the difference between "process efficiency" and "recovery"?

Answer: The term “process efficiency” (PE) as used in this course refers to the joint effect of possible losses during sample preparation and ionization suppression/enhancement (MEionization) in the ion source (see section 5.1). The term “recovery” (R) only refers to possible losses during sample preparation. In most other techniques, this distinction is not made and if recovery determination is carried out then what is actually determined is process efficiency, in the sense that also losses after sample preparation are accounted for. This is OK, if one can assume that the losses during sample preparation are significantly higher than losses during subsequent steps, as is the case in most techniques. In LC-MS this is not so: ionization suppression/enhancement can be a very serious effect. Furthermore, depending on, whether signal loss (usually it is loss, rarely enhancement) occurs via analyte loss during sample preparation or because of ionization suppression/enhancement in the ion source, different approaches are needed for improving the situation. For this reason it is useful to distinguish between R, MEionization and PE and quantify them separately.

Question: What kind of samples are best for control charts?

There are different types of control charts but X charts are addressed here only, as these are the most widespread (and useful).

In the case of X charts different quantities can be plotted on the control chart, depending on what information you want to obtain from the control chart. In most cases people use control charts for keeping the whole analytical method under control. In such a case plotting the result in the control chart is the right thing to do.

Furthermore – it is very strongly recommended not to use analyte in solvent as control sample. Please instead use real sample as control sample and carry out sample preparation every time you run the control sample. Only in such case will the control chart embrace all steps in your method. If you use analyte in solvent then any information about sample preparation (as well as possible selectivity and ionization suppression issues that occur with real samples) will be lost.

A very good source of information on control charts is the Nordtest Trollbook.

Question: How to understand performance characteristics specified eg in Drinking Water Directive (DWD)?

Directive specifies performance characteristics for uncertainty of measurement and limit of quantitation (LoQ). Note, until 31 December 2019 limit of detection (LoD), Trueness and Precision are also allowed to use as performance characteristics. The directive should be read in such a way:

Look for the parametric values in Annex I for a specific compound eg for pesticides 0.10 µg/l (commonly analysed with LC-MS).

Annex III Part B states that limit of quantitation (LoQ) should be 30% or less of the relevant parametric value. Meaning that LoQ for pesticide analysis should be 0.025 µg/l or less.

In Annex III it can be seen that uncertainty of measurement has to be 30% of the parametric value for pesticides with a note that values as high as 80% may be allowed for a number of pesticides.

In the directive the performance criterion for measurement uncertainty (k = 2) is the percentage of the parametric value stated in the table or better. Measurement uncertainty shall be estimated at the level of the parametric value, unless otherwise specified.

However, until 31. December 2019 also the following is allowed: minimum performance characteristics for trueness, precision and limit of detection can be found in Annex III, which for pesticides is 25% for each parameter meaning that in method validation for pesticides, for each compound, trueness, precision and LoD should be equal or better than 0.025 µg/l.

Question: If the blank samples are not available for the type of matrices you are working with  then how can one check for the selectivity of the method?

Absence of blank matrix is relatively common situation, but there is no single solution. Some steps to take.
- Retention time and peak shapes of analyte in sample and standard solution must match. (Comparing unspiked and spiked sample peaks may be helpful.)
- If isotopically labelled standard substance is used, it must also have equal retention time and identical peak shape.
- LC conditions (e.g. different column, eluent pH, gradient profile) should be altered and all the above must still be true.
- In case of MS-detection all the recorded transitions should yield peaks of similar shape and retention time.
- Fragmentation spectra of analyte peak in sample and in standard solution must match.
- ...

Question: What is the benefit of identification points and how they can be calculated?

Answer: Identification points are related to identity confirmation and are used in the EU guidelines 2021/808 [ref 5]. Let's imagine that we register a chromatogram of a urine sample and find a peak at 5 min. From calibration experiments we know that retention time of paracetamol standard substance is also 5 min under the same conditions. Could the peak in sample be due to paracetamol? Yes, but we can not be very sure, because we only used low resolution (LR) MS for m/z 152, but there can be several compounds eluting at 5 min and yielding ions of m/z 152. (1 identification point)
For higher confidence, we could use more sophisticated MS techniques, for example MRM mode to isolate precursor of 152 and fragmentation product of m/z 110. There is relatively limited number of compounds eluting at 5 min, ionizing to m/z 152 and fragmenting to 110 m/z. There is more confidence that the peak is due to paracetamol. (1 + 1.5 = 2.5 identification points)
So, the identification points reflect how much proof we have to claim that the peak is due to specific compound. For calculation examples, please see Example 2 in chapter 2.7 (https://sisu.ut.ee/lcms_method_validation/27-identity-confirmation-examples)