Fresenius J Anal Chem (2001) 371 : 994–1000 DOI 10.1007/s002160101059

O R I G I N A L PA P E R

Christoph von Holst · Anne Müller

Intercomparison study for the determination of selected polychlorinated biphenyls (PCBs) in feed matrices

Received: 15 May 2001 / Revised: 16 July 2001 / Accepted: 19 July 2001 / Published online: 6 October 2001 © Springer-Verlag 2001

Abstract Analytical methodology currently employed for the determination of seven indicator PCBs in three compound feeds and fish meal has been evaluated in a collaborative study. The majority of the obtained relative standard deviations of the PCBs varied from 20 to 30%. On assuming a target relative standard deviation of 22% for the analytical results, statistical evaluation showed that about 80% of the participating laboratories delivered data within an acceptable range of ±44% of the assigned concentration in the test materials. However, between 10 and 20% of the participating laboratories reported unacceptable results. Major problems seemed to arise from insufficient separation of PCB congeners, low extraction efficiency, and calculation errors. Correct identification of the target PCB congeners was a prevalent problem if only one capillary column in combination with an electron capture detector (ECD) was employed. The correct preparation of the calibration solution by the laboratories turned out to be only a minor problem. The laboratories participating in this study employed quite different techniques at all stages of the analytical procedure. Principal component analysis indicated that laboratories using an internal standard tended to report higher values for the target analytes. If the PCB concentrations were related to the fat content of the sample, the variability of the reported results decreased for compound feed but increased for fish meal. These inconsistent results are probably due to the fact that fat is not an objective parameter but is defined by the analytical technique employed. It is assumed that harmonizing analytical methods for the determination of this parameter could improve the precision of the PCB results.

C. von Holst (✉) · A. Müller European Commission, Joint Research Centre, Institute for Health and Consumer Protection, Food Products Unit, I-21020 Ispra (VA), Italy e-mail: [email protected]

Introduction In 1999 the Belgian authorities informed the European Commission that samples of animal feed had been found to contain high concentrations of dioxins [1]. Further investigations revealed that the high PCB concentrations in pork and chicken meat were caused by feed previously contaminated by oil containing PCBs [2]. This finding triggered the need for reliable methods for the determination of PCBs in animal feed. In order to gain an overview of the methodology currently applied in laboratories of EU Member States we organized a collaborative trial in which 27 laboratories from 15 countries participated. While many intercomparison studies for the determination of PCBs in environmental and some food matrices have been carried out in the past, this study is the first one focusing on the analysis of feed. Although many laboratories are experienced in analyzing PCBs, the exact determination of these compounds is still a challenging task, as indicated by the results of the laboratory performance scheme “Quality Assurance of Information for Marine Environmental Monitoring in Europe” (QUASIMEME) [3]. The main task of this program is to assess and to improve the proficiency of laboratories in the analysis of PCBs and other contaminants in marine sediment and biota for monitoring purposes. It revealed a high relative standard deviation (RSD) of the PCB results in mackerel, varying from 30% to 55% depending on the PCB congener, matrix, and concentration of the analyte [4]. The laboratories were also requested to analyze standard solutions, and cleaned and uncleaned extracts in order to track down major reasons for the poor results reported by some laboratories [5]. The authors of this paper emphasized the importance of optimization of GC conditions and the extraction procedure. The objective of the present paper was to elaborate the results of an intercomparison study by assessing the variability of the analytical data reported by the laboratories. In addition, the analytical methods employed for the determination of PCBs in animal feed are discussed and the

995

proficiency of the participating laboratories in reporting reliable results is evaluated. The study was a dedicated collaborative trial for analytical laboratories designated by the EU National Authorities responsible for feed analysis. The participating laboratories were requested to determine the content of seven indicator PCBs (28, 52, 101, 118, 138, 153, and 180) in four feed samples using their own methodology. The laboratories were also asked to analyze a pure iso-octane solution containing the seven indicator PCBs, but the exact concentration of the PCB congeners was unknown to the participating laboratories. As the laboratories used their own calibration standards for the quantification of the PCBs, the variability of the reported PCB concentration in the iso-octane solution was considered as an indicator of the comparability of calibration standards used by European enforcement laboratories. In this study the participants were requested to refer the measured concentration of PCBs to the amount of sample, without taking into account the actual fat content in the material. Maximum levels of PCBs in certain food items expressed on a lipid basis are established by European legislation [1] requiring the determination of the fat content in the sample. Since this approach could also be applied to the animal feed area in future legislation, the laboratories participating in this study were asked to determine the fat content in the test materials as an additional parameter in order to estimate the variability of this parameter.

Table 1 Statistical evaluation of the results from feed 1 and feed 2

Materials and methods Test materials A commercial compound feed for poultry (“Becco Giallo”, Raggio di Sole Mangimi S.p.A., Italy) and fish meal provided by a Belgian feed producer were used as test materials. Fish meal is a typical ingredient for the production of animal feed. The compound feed was fortified with PCBs at different concentrations whereas the fish meal was naturally contaminated with PCBs at about 36 ng (sum of 7 PCBs)/g matrix. This approach allowed us to evaluate the ability of the laboratories to analyze two types of matrices and different PCB congener patterns. The PCB congener patterns evaluated in this study reflect contamination of PCBs originating from technical mixtures such as Clophen A 60 (compound feed) and biological matrices (fish meal). Using technical mixtures of PCBs and naturally contaminated test material also guaranteed that not only the seven indicator PCBs were present in the samples but also other congeners that could interfere with the determination of the target compounds. Three different levels of the sum of the PCBs were chosen, ranging from 6 to 27 ng (sum of 7 PCBs)/g matrix. The low level sample requires measurements close to the limit of quantification whereas the high level sample represents materials which should not pose analytical problems to the majority of analytical laboratories involved in the analysis of PCBs in feed. The “true” concentrations of the PCBs in the test materials were determined from the results submitted by the laboratories by applying robust statistics and are shown in Table 1 and Table 2. Preparation of fish meal and compound feed fortified with PCBs The compound feed was ground using an ultracentrifugal mill (ZM 100 Retsch GmbH, Haan, Germany ) and filled into 210 brown

Feed 1

PCB 28 PCB 52 PCB 101 PCB 118 PCB 138 PCB 153 PCB 180

Table 2 Statistical evaluation of the results from feed 3 and fish meal

Feed 2

No. of labs

Assigned value (ng g–1)

SD (ng g–1)

RSD (%)

No. of labs

Assigned value (ng g–1)

SD (ng g–1)

RSD (%)

21 21 21 21 21 21 20

1.04 1.09 0.97 0.64 0.91 0.85 0.50

0.48 0.29 0.27 0.22 0.25 0.25 0.17

46 26 28 34 28 30 34

22 22 22 22 22 22 22

1.75 1.84 2.04 1.49 2.24 2.00 1.13

0.70 0.55 0.63 0.51 0.71 0.64 0.39

40 30 31 34 32 32 35

Feed 3

PCB 28 PCB 52 PCB 101 PCB 118 PCB 138 PCB 153 PCB 180

Fish meal

No. of labs

Assigned value (ng g–1)

SD (ng g–1)

RSD (%)

No. of labs

Assigned value (ng g–1)

SD (ng g–1)

RSD (%)

23 22 23 23 23 23 22

4.75 4.13 4.48 3.28 4.41 4.02 2.21

1.94 1.17 1.04 0.94 1.35 1.04 0.64

41 28 23 29 31 26 29

20 22 23 23 23 23 23

0.56 1.31 2.76 4.07 9.55 13.5 4.32

0.45 0.54 0.92 1.48 2.94 3.17 1.33

81 41 33 36 31 24 31

996 glass bottles by means of a sample splitter (DR 100 Retsch GmbH, Haan, Germany) and a sample divider (PT 100 Retsch GmbH, Haan, Germany). Each glass bottle contained 20 g of the test material. Three spiking solutions containing different concentration levels of the sum of the target PCBs were prepared by diluting aliquots of technical mixtures (Clophen A 28, 30, 40, 50, and 60, each 1000 µg mL–1 in cyclohexane) in 25 mL of n-hexane. Each spiking solution contained different aliquots of these technical mixtures. For instance, the spiking solution for feed 3 contained 520 µL of Clophen A 28, 1560 µL of Clophen A 50, and 1040 µL of Clophen A 30, A 40, and A 60, respectively. This procedure guaranteed that the concentration level of the individual indicator PCBs in each spiking was of the same order of magnitude but the distribution pattern of the indicator PCBs varied between the spiking solutions. For each feed material (compound feed 1, 2, and 3) 70 bottles were individually fortified with 100 µL of the respective spiking solutions by using a microliter pipette (Eppendorf, Germany). Due to this preparation procedure, the participants of the trial were instructed to analyze the whole content of the glass bottle. Commercially available fish meal with a proven content of PCBs was sieved through a 2 mm sieve to remove foreign matter and the larger sized pieces and subdivided by means of a sample splitter and a sample divider. Each bottle contained 20 g of test material.

PCB concentration level is below 120 ng g–1. This was – the case for each PCB in all matrices and therefore we used 0.22 X for σ in the equation for the calculation of the z-scores. The robust mean of the reported results was also used to define the assigned value of the PCB congeners in the test materials. The z-score was calculated for each PCB in the four test materials, resulting in 28 z-scores for each laboratory. The laboratory performance was considered to be acceptable when the absolute value of the z-score was below 2 and questionable when the absolute value of the z-score was between 2 and 3. The performance of the laboratory was unacceptable when the absolute value of the z-score was above 3 [7]. Multivariate analysis of the results of the laboratories was performed using the software package Unscrambler (Camo ASA, Norway).

Results and discussion Of the 27 laboratories which agreed to participate in the study, 23 reported results. Four laboratories were not able to report results.

PCB standard solution A solution of the seven PCBs in iso-octane was prepared from a concentrated solution of 100 µg PCB congener/mL (Ehrensdorfer GmbH, Germany). The concentration of each PCB congener in the final solution set at 1000 ng mL–1 was unknown to the participants in the trial, but the weights of the ampoules were divulged to the participants to check for losses of solvent during shipment. Homogeneity of the test material Ten glass bottles were randomly selected from each of the four batches of test material. In the case of compound feed the whole content of the bottle was taken for analysis whereas the fish meal was analyzed in duplicate [7]. The homogeneity of the compound feed was evaluated by applying the F-test to compare the received standard deviation of the analytical results of the ten bottles with the standard deviation from the in-house validation of the analytical method [6]. The homogeneity of the fish meal was proved by subjecting the results of the duplicate analyses to one-way “analysis of variance” (ANOVA) as described by Thompson et al. [7], meeting the criteria for sufficient homogeneity as defined in this paper. Statistical evaluation of the results The assigned concentration of the PCBs in the test materials and the corresponding standard deviation were calculated from the reported results of the laboratories using robust statistics. Applying this approach obviates the needs for detection and rejection of outliers, and thus the impact of extreme values on the average and the standard deviation is down-weighted [8]. Moreover, robust statistics works well for distributions that deviate from normal distribution due to extreme values, which is typical for data received in a collaborative trial. In this study the algorithm recommended by the Swiss Food Manual was used [9]. The parameter for evaluation of the performance of the laboratories was the z-score calculated according to the following equation: x− X σ – where X is the robust mean of the reported results of the respective PCB congener, x is the result of the respective laboratory and σ is the target value for the standard deviation. As suggested by Thompson [10] this value should be set at 22% of the assigned value if the z=

Overall performance characteristics of the methods applied In Table 1 and Table 2 the variability of the results reported by the laboratories are presented in terms of standard deviation (SD) and relative standard deviation (RSD). The calculated RSDs for different PCBs varied from 23% to 81%. Comparing the values of the RSD revealed that – with the exception of PCB 28 – the variability of the data did not differ very much amongst the seven PCB congeners. PCB 28 turned out to be the most critical analyte in all matrices, as indicated by the extremely high values for the RSD. If this congener was omitted from the evaluation, the upper range of the obtained RSDs dropped to 41%. No difference between the RSDs of the PCBs in the fortified material and the naturally contaminated test material was observed. Evaluation of the results of the standard solution showed a good correspondence of the robust mean calculated from the participants’ results with the concentration of the PCBs as prepared by the organizer of the study. The RSDs of the PCBs varied from 12% to 16%, reflecting the variability of the calibration solutions used by the laboratories. Since the RSDs of the PCBs in the standard solution are much lower than those in the samples, it can be concluded that the preparation of the calibration solution by the laboratories was not an important error component of the overall variability of the results. A dependence of the obtained RSD on the concentration level of the analyte as predicted by the Horwitz equation [11] was not observed. For instance, the concentration of PCB 138 covered a range from 0.91 ng g–1 (feed 1) to 9.5 ng g–1 (fish meal), but the corresponding RSDs of this analyte in all materials were almost the same (RSD= 28% for PCB 138 in feed 1 and 31% in fish meal). The analytical methods employed by the participating laboratories were quite different. Ten laboratories used the Soxhlet method for extraction of the samples but also cold

997 Fig. 1 Results of principal component analysis. Score plot of the laboratories showing the characteristics of the method applied (E=ECD, M=MS; S=Soxhlet, C=Cold extraction, A=ASE; N=No internal standard; Y=Internal standard; M/C/Y*=Method applied for the homogeneity study)

extraction techniques and modern approaches such as accelerated solvent extraction (ASE) were applied. The solvents chosen for the extraction varied in terms of volatility (n-pentane and toluene) and polarity (n-hexane and acetone). Twelve laboratories added an internal standard to the sample prior to the subsequent extraction using isotopically labeled congeners (five laboratories) or native congeners. In this study, one or more of the following native PCBs were used as internal standard by the individual laboratories: PCB 20, 30, 53, 143, 178, 169, 198, 209. Most often the extracts were treated using silica gel impregnated with sulfuric acid but other solid phase extraction methods based on alumina and florisil were also used. Four laboratories applied gel permeation chromatography and one laboratory used on-line liquid chromatography (LC) coupled to gas chromatography (GC). Capillary columns differing in their polarity (e.g. “DB 1”, “DB 5”) and of length ranging from 25 m to 60 m were used by the laboratories for GC analysis. Most frequently a “DB 5” like column was applied (eleven laboratories) but some laboratories also used special phases such as “DB dioxin”. Four laboratories used two columns of different polarity for unambiguous identification of the target analytes. Thirteen laboratories used mass spectrometry (MS) and ten laboratories employed an ECD for identification and quantification of the PCBs. We also subjected the calculated z-scores of the laboratories to principal component analysis (PCA) in order to establish a possible link between the variability of the results and differences of the analytical methods employed by the laboratories. Two laboratories were identified as outliers since they consistently reported extremely high results shown by z-scores above 5 and were therefore excluded from further statistical analysis. The following methodological characteristics were chosen for statistical analysis: (1) The extraction technique such as Soxhlet, cold extraction, or ASE; (2) the use of an internal standard; and (3) the type of detector – ECD or MS.

By applying PCA the results of the laboratories defined by the original variables (the PCB congeners in the different matrices) are projected onto a smaller number of variables called principal components [12] in order to facilitate the visual interpretation of the relationship between the results of the laboratories and the analytical methods employed. Principal components (PCs) are calculated from the original variables in such a way that the first PC represents the major part of the overall variance of the analytical results [12]. The outcome of PCA presented in Fig. 1 shows the coordinates (scores) of the results of the laboratories in the coordinate system established by PC 1 and PC 2 along with the analytical method applied. Laboratories close to the origin of the coordinate system reported acceptable results whereas laboratories which reported extreme values have high scores along the PCs. Evaluating the outcome of the PCA revealed a clustering of laboratories which used an internal standard in the positive range of PC 1 whereas laboratories without an internal standard tended to have lower scores for PC 1. This also applied to the analytical method employed for the homogeneity study [6] of the test material, indicating a good correspondence with the results of other methods which used an internal standard. The clustering of these laboratories in the positive range of PC 1 means that the average of the reported PCB concentrations of the laboratories with an internal standard were higher than those from laboratories without an internal standard. We suggest that in this study a methodological bias due to the use of an internal standard contributed to the overall variability of the results. The grouping of the laboratories in respect of the use of internal standard correlated well with the type of detector (ECD or MS) employed. This correlation is probably due to the fact that the higher selectivity of mass spectrometry compared to ECD allows the easier use of suitable internal standards. Moreover, the outcome of the principal component analysis did not hint at a major influence of other methodolog-

998 Table 3 Performance profile of the laboratories (feed 1 and feed 2)

PCB 28 PCB 52 PCB 101 PCB 118 PCB 138 PCB 153 PCB 180

Acceptable range (%)

Questionable range (%)

Unacceptable range (%)

Feed 1 Feed 2

Feed 1 Feed 2

Feed 1 Feed 2

76 90 81 81 86 90 85

5 5 10 10 10 0 5

19 5 10 10 5 10 10

73 83 86 77 77 73 77

9 9 5 9 0 18 5

18 9 9 14 23 9 18

Table 4 Performance profile of the laboratories (feed 3 and fish meal)

PCB 28 PCB 52 PCB 101 PCB 118 PCB 138 PCB 153 PCB 180

Acceptable range (%)

Questionable range (%)

Unacceptable range (%)

Feed 3 Fish meal

Feed 3 Fish meal

Feed 3 Fish meal

78 82 87 83 78 83 82

9 9 0 0 9 4 5

13 9 13 17 13 13 14

50 68 70 78 83 83 83

10 18 17 4 4 13 4

40 14 13 17 13 4 13

ical factors such as the type of the extraction method on the distribution of the analytical results of the laboratories. Overall performance characteristics of the participating laboratories Table 3 and Table 4 show the percentages of acceptable, questionable, and unacceptable laboratories for each PCB Fig. 2 z-Scores of the laboratories for PCB 138 in fish meal (assigned value: 9.6 ng g–1)

congener based on the z-scores calculated for the laboratories. Applying the acceptability criteria explained above allowed interpretation of the outcome of this evaluation in a very simple way as we used the same target relative standard deviation of 22% for each PCB in the four test materials. As an example, the z-scores of the laboratories for PCB 138 in fish meal shown in Fig. 2 indicated that 83% of the laboratories reported acceptable results. The assigned value of this congener was 9.6 ng g–1 and the corresponding target standard deviation was 2.1 ng g–1, which corresponds to a RSD of 22%. The acceptable deviation of the reported concentration from the assigned value was therefore ±4.2 ng g–1 (2×2.1 ng g–1). If the results of PCB 28 were omitted, the percentage of acceptable laboratories varied from 68% to 90%, depending of the PCB congener under investigation. Between 10% and 20% of the laboratories reported results that were far beyond the acceptable range. Careful scrutiny of the results and the chromatograms of these laboratories allowed some tentative reasons to be put forward. For instance, some laboratories reported extremely high values for the PCB concentration in the samples but the corresponding results of the analysis of the standard solution were in an acceptable range. Examination of the chromatograms of these laboratories confirmed good separation of the target analytes and therefore we assume other reasons for these discrepancies, such as calculation errors. In some cases, insufficient separation of PCB congeners (e.g. PCB 28 from PCB 31) might contribute to the overall variability of the data. This problem seemed to be prevalent if an ECD with only one capillary column was employed. de Boer reported on the evaluation of another collaborative trial for the determination of PCBs in fish and suggested the use of two columns of different polarity with a minimum length of 50 m and maximum internal diameter of 0.25 mm for sufficient separation of the target analytes [3]. Some results hint that PCB 138 might have been mistaken for PCB 153, which can occur because in

999

most columns PCB 153 elutes before PCB 138, but the order of reporting the results in the spreadsheet was the other way round. Some laboratories reported low values for the PCB concentration and the fat content indicating a non-quantitative extraction. The above discussion concerned the results of the individual PCB congeners. We also calculated the sum of the seven PCBs from the concentrations of the individual congeners reported by the laboratories. The RSD of this parameter was 26% for feed 1, 25% for feed 2, 24% for feed 3, and 26% for fish meal. The results indicated that the variability of the sum of the congener was smaller than the variability of the individual congeners. This finding also demonstrated that the performance criteria for the employed analytical methods depend on the purpose of a specific exercise. For instance, methods suitable for monitoring purposes need to measure all analytes with sufficient precision at a presumably low concentration level in order to spot a yearly trend of this concentration. However, methods used in emergency cases to screen samples for exceeding the threshold levels of the sum of the PCBs do not have to be extremely precise regarding the determination of each individual PCB. A good example is PCB 28, which is of less importance if the source of the contamination was highly chlorinated technical mixtures such as Chlophen A 60. Determination of fat in the sample Seventeen laboratories reported results obtained for the fat content whereas six laboratories omitted the fat determination, as this parameter was not included in their normal analytical program. We calculated the robust mean and the corresponding standard deviation of the fat content in fish meal and the compound feed. In this calculation the obtained values of all three compound feed materials were combined as they are originated from the same batch. The assigned value for fish meal was 11.3% (RSD=11%) and 3.8% for the compound feed (RSD=13%). However, we did not assign a z-score of this parameter to the laboratories as the fat determination is an empirical method and what is actually measured is defined by the method employed [14]. This problem has also been addressed in another paper showing that the extractability of polar phospholipids from biological matrices depends on the characteristics of the solvent used for the extraction [13]. In order to examine the influence of the determination of this parameter on the reported PCB concentrations, the variability of the data based on the matrix and on the measured fat content were compared. The comparison shown in Table 5 revealed that for feed 1, 2, and 3 the greater part of the RSDs were lower when the PCB concentrations were expressed on a lipid basis, whereas the opposite held true for fish meal. Though the data of this study do not allow a clear explanation, the differences between the test materials in the concentration of the fat and the classes of lipids involved could have contributed to these contradictory findings. Since the use

Table 5 Difference of the calculated RSDs of the PCBs based on the matrix and on the fat content RSD related to the matrix minus RSD related to fat equals the figures below

PCB 28 PCB 52 PCB 101 PCB 118 PCB 138 PCB 153 PCB 180

Feed 1

Feed 2

Feed 3

Fish meal

–6 12 12 0 13 3 7

1 5 11 1 15 13 5

2 4 4 1 2 2 –1

7 –7 –3 –12 –2 6 –3

of different methods for the determination of the fat content could introduce an additional error in the variability of the PCB concentration when calculated on a lipid basis, the outcome of this comparison also indicates a need for harmonization of such methods, as proposed in the ISO method 6492 [15].

Conclusion Evaluation of the results of the interlaboratory study for the determination of PCBs in feed showed that about 80% of the participating laboratories were able to report acceptable results based on a target relative standard deviation of 22%. However, between 10 and 20% of the laboratories seemed to have severe problems regarding the determination of PCBs in biological matrices. The variability of the PCBs in the standard solution measured by the participating laboratories did not significantly contribute to the overall variability of the PCBs in the test materials. If the measured concentration has to be expressed on a lipid basis due to legal requirements, harmonization of the fat determination is recommended. We assume that the quality of the results can be improved by imposing more efficient quality assurance systems on the laboratories. If new analytical methods are implemented in the laboratories, in-house validation procedures should be applied. The main components of such a procedure are (1) determination of performance characteristics, (2) ruggedness test, and (3) traceability of the results using certified reference material. However, it should be pointed out that the currently available certified materials such as fish oil or milk powder are not perfectly suited for this application, indicating a strong need for the production of such material in the field of contaminants in feed. Acknowledgements The authors are grateful to the following laboratories for their participation in the study. Institut für Pflanzenschutzmittelprüfung, Wien, Austria; Bundesamt für Agrarbiologie, Linz, Austria, Umweltbundesamt, Wien, Austria; Laboratorium Tervuren, Tervuren, Belgium; Laboratorium Ghent, Gent, Belgium; Laboratoire de Liège, Liège, Belgium; Plantedirektoratet, Lyngby, Denmark; Laboratoire Interrégional DGCCRF de Rennes, Rennes, France; Laboratoire Interrégional DGCCRF de Lille, Vil-

1000 leneuve d’Ascq, France; AFSSA Paris, Maisons Alfort Cedex, France; LUFA Kiel, Kiel; Germany; LUFA Speyer, Speyer, Germany; LUFA Hameln, Hameln, Germany, Landesanstalt fuer Landwirtschaftliche Chemie, Hohenheim, Germany; Institute for Food Hygiene of Athens, Athens, Greece; State Laboratory, Dublin, Irland; Istituto Zooprofilattico Sperimentale della Lombardia ed Emilia Romagna, Brescia, Italy; Istituto Zooprofilattico Sperimentale dell’Abruzzo e Molise “G. Caporale”, Teramo, Italy; RIKILT-DLO, Wageningen, Netherlands; Norwegian Institute for Air Research – NILU, Kjeller, Norway; Laboratorio de Espectrometría de Masas, Barcelona, Spain; Laboratorio de Sanidad Animal, Algete, Spain; Instituto Químico de Sarriá, Barcelona, Spain; Alcontrol Nyköping, Nyköping, Sweden; Laboratory of the Government Chemist, Teddington Middlesex, United Kingdom

References 1. European Commission (1999) Commission Decision 1999/ 788/EC Official Journal of the European Commission L 310: 62–70 2. Bernard A, Hermans C, Broeckaert F, De Porter G, De Cock A, Housins G (1999) Nature 401:231–232

3. de Boer J, Wells D (1996) Mar Poll Bull 32:654–666 4. de Boer J, Wells D (1997) Mar Poll Bull 35:52–63 5. de Boer J, van der Meer J, Brinkmann UA (1996) J AOAC Int 79:83–96 6. von Holst C, Müller A, Björklund E, Anklam E (2001) Eur Food Res Technol 213:154–160 7. Thomspson M, Wood R (1993) J AOAC Int 76:926–941 8. Thompson M, Lowthian PJ (1996) Analyst 121:1597–1602 9. SLB, 1989 (1989) Statistik und Ringversuche. In: Schweizerisches Lebensmittelhandbuch. Kapitel 60. Eidgenössische Drucksachen- und Materialzentrale, Bern, Switzerland 10. Thomspon M (2000) Analyst 125:385–386 11. Horwitz W, Albert R (1996) J AOAC Int 79:589–621 12. Esbensen KH (2000) Multivariate data analysis – in practice. Camo ASA, Oslo 13. Ryan JJ, Mills P (1997) Chemosphere 34:999–1009 14. Eurachem/Citac Guide (2000): Quantifying Uncertainty in Analytical Measurement available via internet: http://www.measurementuncertainty.org/ 15. ISO/FDIS 6492 (1999): International Organization for Standardisation, Genève, Switzerland