The usual criteria for rejection are when a result is outside the 3 certified standard deviation of the CRM. Under the assumption that the CRM should follow a normal distribution about the certified mean (with the certified SD), then anything outside 3SD is a rare event (about 3/1000 by chance should occur this way). I've seen many people use 2SD as a rejection limit but this is wasting your money as 1:20 CRMs should return values outside 2SD. However, you should also track trends in the data and watch out for consistent bias high or low in the CRM. Some use of the standard error for multiple results will help to test whether a bias is significant. It is also worthwhile checking that the results are in fact normally distributed - especially for some the low concentration variable in nickel laterites.

Adding to above comments, if CRMs samples consistently track outside of +/-2SD it might be worth picking up the phone and chatting to the lab. This assumes matrix matched standards have been used.

For duplicates, you can measure paired CV values against an 'acceptable' error distribution (which depends on the application) or compare against stable historic error levels.

Thanks for your response. We will use these criteria, and when we have enough data, we will review trends.

It is also important to remember to look at ALL standards in the batch, if only one standard out of 5 say is more than 3 SD's from expected, is this necessarily reason to reject a batch?? I err on the side of if more than half of the standards fail then definitively reject the batch, otherwise chat with the lab to try and work out why only one or two standards have failed.

One should look beyond one standard failing (that is, reporting either below or above the expected limits). Errors such as transcripts or topographical, standard swaps either during insertion in the field or lab are common human errors very difficult to eliminate especially when more than one standard types are utilized for a batch. However, few checks are put in place to control those human errors in some establishment (see me for some tips). There are several checks or things to take into consideration especially by yourself before you bounce on the lab for standards failure. We normally overlook the field problems or mistakes during standards insertion especially, but always shift blame to the lab.

CRMs inserted need to be within the ranges of the AAS instrument calibration range; laboratory specific used to analyze the samples, otherwise the results will be meaningless. For tracking trends the SPC Nelson(1984) 8 rules should be applied to detect bias. The SD used after at least 20, preferably 30 or more assays of a CRM should be applied to establish the precision of the laboratory, the SD accompanying the CRM certificate can only be used as a guide in the early stages and should not be used. A good laboratory will achieve close to the certificate SD. Also a batch submitted to the laboratory should match the laboratory fusion if using FA or digestion batch otherwise a consignment submitted may be composed of many laboratory fusion batches.

You have highlighted a big problem when the laboratory is under pressure, but a bigger issue is to ensure each sample submission to the laboratory is correct, legible preferably electronic spreadsheet (accompanying hard copy) and signed off before leaving the site to minimize finger pointing and embarrassment later on.

MMCRMs whilst good may not reflect the performance of the routine samples considering the wide variation in sample matrix. Comparability tests should be performed on at least 5% of the mineralized samples, preferably at a recognized superior laboratory using a superior analytical method such as part & weigh for gold by FA.

One should remember that the company submitting the samples and the laboratory receiving the samples is handled by the least skilled of both organizations and least equipped to sort out any issues.

Thanks for your comments; they are very useful, especially at this stage of implementation of the QA/QC program.

I have a question, the results obtained show that CRMs are between 2 and 3 SD and internal laboratory bias is always negative (underestimated). This is normal when CRMs are commercial (not MCRM)?

Depends on the magnitude ofthe bias and whether the bias is 'statistically significant'. To do this run soe t-tests on each analyte (see http://is.gd/qXXscY as an example ). Note you should check SDs is about the same or else you may be having a precision issue - here your testing for mean difference under the assumption the precision is comparable.

My experience is a commonresult is that CRMs usually show some small bias (perhaps within +/- 0.5 the certified SD) with a primary laboratory - and thisis to be expected. If you look at the certification certificate you should findthat the certified values (mean and SD) are average of the results from anumber of reputable laboratories around the world (or country) - the idea hereis that the certified values will vary not only due to the heterogeneity of theCRM material but the different sampling errors (hopefully small) occurring ateach laboratory. The other issue can be the analysis methodology.

First read the CRM certificate and make sure that the analytical method used to certify the CRM is the same as what you are asking your laboratory to do. I've seen CRMs certifiedusing more than analysis one method for example XRF and ICPOES and the certifiedvalue is the average of different methods. If you are using XRF than perhaps you should only use the CRM results that did XRF - this may remove the bias effect. Recalculate the CRM values and see if this removes the bias - document what you have done for audit purposes.

You should also check ifthere is a 'outlier' laboratory in the CRM certificate that has reported values much higher or lower than the other labs. Ideally the results from all labs should be an approximately normal and if there is one laboratory introducing askew in the data - removing it may be justified. Again remove and recalculate the CRM values and so on.

Additionally I would recommend is sending at least 20 of your CRM packets - along with rejects from original samples to one of the reputable umpire laboratories on the certification certificate - perhaps 1 CRM to every 10 check samples. You should be doing this anyway on some of your samples as a quality control check. If the results for this CRM are biased away from the certification lab results on the CRM certificate you may have an issue with your laboratory having a calibration issue. In this case you may wish to talk to your laboratory manager to check machine calibrations and so on.

Finally, the bias you observe may be just the fact that you laboratory will report the CRM a bit lower than the average of the CRM certification laboratories. You may just decide this is where your lab sits in the certification zoo and accept that the running mean (an SD) to date are the values you should use to monitor accuracy.

Just reset the CRM to the average-to-date values and document your reasons.

Be aware of the dangers of chasing noise when assessing bias using multiple t tests, particularly when looking at multiple parameters. Use of Hotelling's multivariate T squared test is suggested.

Thanks for that advice.

Referring back to the comment that has been made, I'm interested in knowing the reasons why one would use the standard deviation of the batch rather than the standard deviation on the CRM certificate. I have heard a number of people say that you should use the value on the certificate and others say you should use the value you derive from the laboratories results. What would be the reason for differences in the standard deviation if the laboratory is using the same analysis technique as that which was used to derive the certified results (reasons other than laboratory hygiene)?

Yes there is a lot of confusion to use or not to use the SD accompanying the CRM certificate. Frequently QA papers are silent on which SD to use and common even for "reputable" consulting companies also to recommend using the certificate SD. Be aware of CPs or QPs acting for banks. Using the Certificate SD will indicate a tighter UCL & LCL than necessary for SPC but for a good laboratory, both SD should be close, but you need the evidence to demonstrate it.

In Explore No 118 p9-13 in a technical note Barry Smee raised this issue for clarification and commented on by Lynda Bloom, Leaver (CCRMP), Hamlyn (Ore Research &Expl PL) and Malcolm Smith (Rocklabs).

The certificate Confidence interval is a statement of accuracy and the SD (or CV) is a statement of homogeneity of the CRM, the results obtained by repeat analysis by a laboratory and between laboratories as part of the proficiency testing according to ISO Guide 35 (1985).

Repeat analysis of the inserted CRM by a user laboratory, establishes the precision achieved by that laboratory for that method. Precision controls the size of bias we can detect; we cannot detect a bias that is smallerthan the SD of the results. So the SD established by at least 20 repeat values (SGS - Rocks to results - a practical guide to laboratory operations), preferably 30 or more to statistically establish the precision limits and UCL & LCL based on 3SD. Any less than 20 repeat analyses of a CRM cannot statistically establish whether a batch can be accepted or not and the SD on the certificate can be used as a guide until at least 20 repeat analyses are available. Earlier indicates the "RM is used to determinewhat the laboratory actually achieves, not whatit should achieve."

CRM's are typically analyzed by multiple labs with variable capability with multiple techniques. For many certificates look at the lab technique employed and match only that data with YOUR labs technique for a guide to precision. As such the stated SD on certificate is generally greater than it will be within a single lab. CRM is to assess accuracy. Your Labs should be able to make a statement on their precision capability based on control charts and basis for rejecting batches.

Thanks for your comments. Reading found the following: "The errors in assays results are usually normally distributed. This means that the expected standard deviation can be used to predict the probability of CRM results falling within particular grade limits or thresholds:

68% of results should fall within +o- 1 sd of the expected grade.
90% of results should fall within +o- 1.65 sd of the expected grade.
95% of results should fall within +o- 2 sd of the expected grade.
99.7% of results should fall within +o- 3 sd of the expected grade. "