In meat fat a build-up of PCB's and dioxins from the environment has been observed. The principal ways by which PCB's are coming into the environment are: vaporisation from plasticisers, combustion of industrials and domestic waste, transformers and capacitors cooling liquids (according to data of the World Health Organisation). The nature and dynamics of the distribution persistence, chemical inertness, spreading in the atmosphere due to fairly high vapour pressure and capacity for diffusion in the soil and surface waters combined with lipophilicity promote the bioaccumulation of PCB's in the fatty tissues of man and animals in amounts considered to be dangerous for vital activity. This is the reason why a lot of governmental organisations in Europe have launched guidelines for foodstuffs, which may not contain more then 200 ppb PCB's per 100 gram fat. The need for a quick and easy screening method has never been bigger nowadays.

Moreover, for dioxin measurement, the relation with PCB is commonly accepted. The PCB level is 1000 times the dioxin level. The measurement of PCB's is correlated to the dioxin level and is a way to measure the dioxin level indirectly.

The common analytical method for PCB measurement in fat is Gas Chromatography (GC). The fat content is usually first extracted from the fat tissue. This extraction is further saponified and the organic fraction is used to inject in the gas chromatography analyser.

Gas chromatography is a separation technique. the samples are separated in different fractions. The different

components coming out of a chromatography column are detected. Different detection techniques can be used. Like mass spectroscopy (MS) or electron capture detector (ECD).

In the latter gas chromatography detection technique is used, because halogens are very well detected by this technique and PCB's contain a lot of chlorinated bondings.

For PCB a family of 7 congenating piques, each pique of approximately 20 ppb level for the 200 ppb level samples, can be detected and defined as the PCB concentration level.

The standard deviation read-out is approximately 20%.

Although these techniques are providing good sensitivities of the measurements, they require highly skilled personnel, expensive hardware and they are time consuming.

The measurement of components, other than PCB in pork fat, like moisture, fat, fatty acids, saturated and unsaturated fat content, with NIR spectroscopy is widely spread, because of the easiness of the measurement and the analysis time, usually a few seconds to a minute. The measurements do not require highly skilled personnel and NIR spectroscopy is a non-destructive measurement. No reagents are needed and sample preparation is very limited. However for NIR measurements, the practical minimum detection level according to prior art is above 0, concentration. For the measurement of low concentrations, in ppm level, NIR is not used, because in the spectra due to low concentrations in the sample, nearly undetectable changes occur.

The object of the present invention is to provide a process having a superior combination of good sensitivity and easy handling with respect to prior art.

This is achieved according to the invention by a process for determination of halogenated organic compounds, especially PCBs in at least partially hydrophobic samples like fat, using photometry, especially absorption spectroscopy, wherein chemometric calibration models

calculated from spectral data at several wavelengths in the near infrared wavelength range, preferably between 1100 and 2500 nm, from calibration samples with various concentrations of the compound (s) to be determined are used for the determination of the compound (s) in a sample by calculating concentration characteristics from their spectral data at said wavelengths with said calibration factors. NIR-spectroscopy is an analytical method. Results are calculated from reference data, e. g. obtained from different analytical procedures like gas chromatography, using statistical algorithms. This means that known samples are used to calibrate the NIR instrument. NIR ranges from about 750 nm to 2500 nm. At least one chemometric calibration model is calculated using factor analysis and/or regression techniques are used and later on used to predict unknown samples. Calibration results are calculated from reference data obtained from different analytical procedures, e. g. gas chromatography, using statistical algorithms. This is called the calibration process. The unusualness of this invention is the use of NIR for the detection of components in ppb (parts per billion) level.

This means for PCB in pork fat a level of approximately 200 nanogram per 100 gram fat. This is extremely low and thought to be beyond the NIR detection limits. These calibration factors can be determined with mathematical algorithms like those implemented in the SESAME software, commercialised by Bran+Luebbe GmbH (Werkstr. 4 D-22844 Norderstedt/Germany). Preferably the spectral data are recorded advantageously within a wavelength range between 1100 and 2500 nm, where characteristic near-infrared signals, e. g. of PCB are found and where standard near- infrared spectrometers can be used.

The selection of calibration sample concentrations comprises usefully increasing differences of successive samples with increasing concentrations, especially with a concentration range extending up to more than 10-fold of

the minimum detection level of the process to provide good calibrations in the lower concentration level of the calibration samples and/or in order to have much variance represented by the calibration samples. Preferably the selection of samples have a concentration range extending up to more than 10-fold of the minimum detection level.

Especially mixtures of 0,140,280,1400 and 2800 ppb PCB per 100 gram blank fat are used as calibration samples, in particular in combination with real fat samples comprising PCB concentrations within this range.

Transmission and/or reflectance data and/or absorbance and/or the first and/or the corresponding first or higher derivation of these data may be recorded in the process according to the present invention for the measurements.

This can be done by conversion of the measured data according to known techniques, as well.

Surprisingly good correlations are achieved, when for the determination of the compound (s) spectral data of the samples are taken at one or more of following selected wavelengths 17305nm, 17405nm, 22605nm, 22645nm, 22585nm, 13765nm, 13845nm, 21845nm, 13765nm, 1384+5nm, 2180+5nm.

An advantageous embodiment of the process of the present invention comprises a selection of preferred wavelengths is used for recording spectral data, especially the combination of at least three wavelengths, preferably containing 17305, 1740+5 and 22605nm, or 17305, 1740+5 and 22645nm, or 1730+5, 1740+5 and 22585nm, or 1376+5, 1384+5 and 21845nm, or 13765, 13845 and 21805nm in order to reduce measurement time and/or to obtain a good accuracy.

In another preferred embodiment of the process only a subrange of NIR, including a number of selected wavelengths, is used. A step size between 1 and 10 nm,

preferably 2 nm is used in order to capture and represent much variance in the spectral data by the determined calibration factors. For example the selection algorithms, implemented in the SESAME software, commercialised by Bran+Luebbe GmbH, can be used for this purpose.

Preferably the chemometric model used is multiple linear regression (MLR) and/or partial least squares regression (PLSR) and/or principal components regression or cluster analysis. These powerful kinds of algorithms are for example provided with the SESAME software, commercialised by Bran+Luebbe GmbH, and described in the corresponding manual.

In a preferred embodiment of the process of the present invention the determination of the calibration factors comprises a cross validation of calibration samples and/or wavelengths. In this case the best calibration samples will be selected by subsequent removal of calibration data of a few defined calibration samples and verification of the corresponding calibration results. This can be carried out by using the corresponding functions of the SESAME software, commercialised by Bran+Luebbe GmbH. Preferably the calculations and/or the selections are performed at by a software, comprising operations controlling a spectrometer used in the process, wherein calibration data are saved on a computer, possibly via downloading from a data network connected to said computer. For this purpose any suitable combination of standard computers like a personal computer and software products like the SESAME software, commercialised by Bran+Luebbe GmbH and possibly internet software tools can be used. In the latter case calibration data then have to be recorded and calculated only once for several experiments or recording of them may be completely omitted by simply downloading them, e. g. from a data-network link to the manufacturer of the spectrometer used in the process of the present invention, especially

via Internet. Common suitable data transfer techniques may be used for this purpose.

Preferably the samples are liquefied via melting, especially by heating in a microwave and/or chemical reactions and/or enzymatic reactions. Especially fat tissue samples, available in solid and measurable in liquid state will be treated this way. Any suitable enzyme for degradation of biological samples like lipases, proteases may be employed in this process.

In another embodiment of the process of the present invention the sample is an extraction from an initial sample with organic solvents. This extraction from an initial sample with organic solvents enables analysis of various initial samples and decreases the content of moisture and corresponding near-infrared signals in the samples to be measured. Also water samples can be analysed with this process. Any kind of suitable extraction like Soxhlet or shaking an aqueous liquid with organic solvent can be employed for this purpose.

In a further preferred embodiment of the process of the present invention mixtures of 0,140,280,1400 and 2800 ppb PCB per 100 gram blank, that means PCB-free, pork fat are used as calibration samples, in particular in combination with real fat samples comprising PCB concentrations within this range.

Preferably a reflectance recording spectrometer is used for a process according to the present invention, which comprises at least one lead sulphide detector for sensitive signal detection and preferably at least one diffusely gold-plated reflector disposed on the bottom of a sample cup, preferably comprising an electric resistor under the bottom of a sample cup, especially in a cavity within the bottom of a sample cup, preferably in combination with a temperature sensor, preferably a Peltier element directly

fitted under the sample cup and connected to a heat controller which is further connected to the resistor for temperature control. For example the Bran + Luebbe InfraAlyzer 500 can be used as such kind of spectrometer.

This device enables, e. g. to maintain the liquid state of a melted fat sample and/or good temperature control.

An advantageous application of a process and/or spectrometer according to the present invention is detecting PCB in fat on a level between 200 and 1000 ppb per 100g fat, e. g. for surveillance of food quality in accordance with proposed or prescribed limits for PCB contamination.

Another advantageous application of a process and/or spectrometer is detecting halogenated organic compounds in biological samples like food stuffs or in liquefied polymers, petrol or other organic liquids like gasoline and nafta based products, the combustion of which generates extremely toxic substances when halogenated organic compounds are present, or in water like fresh water supplies.

A further advantageous application of a process and/or spectrometer is detecting dioxin indirectly since PCB measurements provide an indirect measure for dioxin as described above.

The following examples will illustrate the invention without restricting it to these examples. Some features will be described referring to the drawing in which same elements are referenced with the same references: Figure 1 shows a cross section through a sample bearing connected to a temperature controller, all used in a spectrometer according to the present invention.

The sample bearing in Figure 1 comprises a sample cup 1 namely a Bran+Luebbe Transflectance Cup (type Dutch cup), which has a diffusely gold plated reflector 2 as a bottom and a path length of 0.2 mm + 0. 01mm. A hole is drilled under the reflector 2, until just 1 mm under the gold plated surface and a PT100 resistor 5 (type DM503) with a positive temperature coefficient of 100 Q at 20°C is mounted in this drilled hole and fixed with a specific temperature paste: heat sink compound (type 340 Dow Corning). The resistor is connected to a temperature controller 6 with cables 7 (type: 17-8811-46342300, from Bartec), outside the cup, which is then further connected with cables 7 to a Peltier element 3 (Thermoelectric Module type DT6-6LS, from Marlow Industries), glued under the sample cup 1. The temperature controller 6 is connected to a outside power supply (type EA-3003 S-Current, 2.5 A, not shown). The sample cup 1 is covered with a quartz window 3 covering the sample space 4 and the sample cup 1 is placed in the Solid Drawer of a Bran+Luebbe InfraAlyzer 500 spectrometer (not shown), during the complete measurement.

Only the quartz window is opened to apply the sample.

The NIR Analyser used for this experiment, is a Bran+Luebbe InfraAlyzer 500 X, equipped with the Standard Solid Drawer.

The wavelength range is 1100 to 2500 nm, selected with a holographic grating. The selected step size, was 2 nm for this experiment to give 701 steps for each spectrum. The used detectors are a pair of lead sulphide detectors and a

diffusely gold-plated integrating sphere is used to collect the reflected light by the sample.

This spectrometer described above was connected to a Personal Computer with the SESAME software commercialised by Bran+Luebbe GmbH, which was used for all data conversions and calculations.

For measurements the temperature can be adjusted by the temperature controller and has been fixed at 50°C +/-1°C, the optimal temperature for fat measurement, so that fats were in a liquid phase, without crystallisation, at this temperature.

A selection of 8 pork fat samples is used, the concentrations of which, between 70 and 2695 ppb PCB per 100 g sample, have been determined by gas chromatography detection. The samples were heated in a microwave for 4 minutes, and the fat substance coming out of the pork tissue, is collected, approximately 5 ml. A sample without PCB (blanco) and a PCB standard sample of 20.000 microgram per ml in iso-octane solution, is then used for making a concentration gradient of five samples, with different PCB concentrations. The standard has first been diluted to a solution of 1 picogram in hexane, and used to spike the blanco sample. Samples with the following concentrations have been made: 0,140,280,1400 and 2800 ppb PCB per 100 gram blank fat. Also these"spiked"samples are used for the further calibration. A list of all sample measurements and sample concentrations is given in table 1.

The pork tissue samples have been heated in a dish, one after the other in a common household microwave until melted, for 4 minutes. The collected fat in the dish has been used for the measurement. To be sure, that the solvent in the spiked samples is evaporated, these samples were also heated in the microwave. To avoid the influence of the solvent in the samples for the calibration a second heating

in the microwave took place. For each sample the collected melt fat was poured into a small lab beaker. This microwave heating was also applied to be sure that less water is remained in the samples.

The melted samples were applied on the gold reflector of the sample cup 5, by a pipette. The samples have been carefully sucked in a pipette, avoiding solid particles and air bubbles. Approximately 4 drops were sufficient to apply a layer of 0.2 mm on the reflector. The cup is then closed with the Quartz window. Some time is needed to be sure that the sample has reached the temperature of 50°C in the cup.

An additional time period of 30 sec. is used to be sure the temperature is not varying. Then the measurement (scanning) is started. After the measurement the cup is opened and cleaned with a paper tissue.

First the spiked samples have been measured in duple. Then the other real samples have been measured also in duple.

All these samples were measured in random order like shown in table 1, to avoid false correlations with temperature, etc.. After this measurement cycle some samples, spiked and not spiked samples, have been measured a third time, to be sure no other changes during this experiment took place.

Though here are used liquid samples, the invention may also be used for analysing solid samples.

Table 1: used samples for the experiment: Spectrum # PPB PCB in fat *1: 0 *2: 0 *3: 2800 *4: 2800 *5: 1400 *6: 1400 *7: 0 *8: 0 *9: 1400 *10: 1400 *11: 2800 *12: 280 *13: 280 *14: 140 *15: 140 16: 70 17: 70 18: 184 19: 184 20: 1046 21: 1046 22: 719 23: 719 24 : 345 25: 345 26: 650 27: 650 28: 1315 29: 1315 30: 2695 31: 2695 32 : 345 Minimum : 0. 00000 Maximum: 2800 Mean : 913. 53

# Spectrum: identification number, given to the measured spectrum ppb PCB in fat: level of PCB per 100 gram fat (*) spiked samples = samples with standard PCB added to a blank fat without PCB

Conversion and calibration calculation was performed in the measurement data of the spectra listed in table 1 and used for prediction of the PCB concentration properties of these measurements according to the following examples: Example 1 : cluster analysis SESAME software, version 3.1 commercialised by Bran+Luebbe GmbH was used in this example. On the measured spectral data a conversion from reflectance data to first derivative data has been applied and the as qualitative analysis approach of cluster analysis was used as calibration calculation. Two of the spectra were determined to have excessive deviations and were deselected by the software from the set of calibration data. The selected calibration spectra were divided into two classes, one of which for 22 spectra from samples with more than 200 ppb PCB per 100g fat and another one for 8 spectra from samples with more than 200 ppb PCB per 100g fat. According to the algorithm of the software, which is based on a principal component analysis (vectorial) calibration factors have been calculated which describe the properties of the two classes of spectral data. From these calibration factors the so called scores factor 2 generated by the software has been selected.

Figure 2 shows a plot of the scores factor 2 against itself for better visualization, resulting in the representation of each of the selected calibration spectra factors as circles in this plot. Two clusters of scores factor 2 values were clearly distinguishable observable as two clusters of overlapping circles in figure 2. The cluster in the upper right area of the plot corresponds to spectra from samples below 200 ppb PCB per 100g fat and the other cluster extending to the left edge of the plot corresponds to spectra from samples over 200 ppb PCB per 100g fat.

This indicates that the differences between the samples are obvious. Based on these calculated calibration factors reconstructions of the initial spectra were calculated from the corresponding concentration values, the deviation of which is represented by residual values. The data are given in table 2.

Table 2: cluster analysis No of series: 2 series, 32 species No. of data points: 701 Wavelength range 1100.. 2500nm, 701,2 nm steps Calibration method: Cluster Calibration set: 2 series, 30 spectra Test set: 0 series, 0 spectra Not selected: 0 series, 2 spectra Wavelength range 1100.. 2500nm, 701,2 nm steps Transformation: First Derivative -First Derivative Product code digits: 52) Rank: 233 Selected Factors: 23) Limits of radii: calculated Limits of residuals: calculated Max. allowed 0.000006 residual' : REGRESSION RESULTS Cluster List of library substance selected spectra name 1. Spectra: 1-22 >200 (*) 2. Spectra: 1-8 <200 (*) List of classes name Number of spectra components5) 1. >200 (*) 22 1 2. <200 (*) 8 1

Table 2: cluster analysis (continued) Calibration set validation Spectrum name# class component5) radius6) 1 1 1 0.085718 2 1 1 0.087269 3 1 1 0.031491 4 1 1 0.032620 5 1 1 0.068760 6 1 1 0.066104 7 1 1 0.098205 8 1 1 0.178520 9 1 1 0.171879 10 1 1 0.088600 11 1 1 0.081277 12 1 1 0.026644 13 1 1 0.042317 14 1 1 0.111513 15 1 1 0.092878 16 1 1 0.105153 17 1 1 0.111212 18 1 1 0.079952 19 1 1 0.083558 20 1 1 0.353181 21 1 1 0.350841 22 1 1 0.036039 23 2 1 0.045488 24 2 1 0.044408 25 2 1 0.043542 26 2 1 0.044338 27 2 1 0.026644 28 2 1 0.030000 29 2 1 0.043125 30 2 1 0.037048 Validation for all spectra Calibration set: 30 Spectra with average residual 7.9e-007 Test set of which 0 spectra

Table 2: cluster analysis (continued) # real predicted Residual identity identity 1 > 200 (*) 1 3e-008 2 > 200 (*) 1 6e-008 3 > 200 (*) 1 4e-007 4 > 200 (*) 1 5e-007 5 > 200 (*) 1 2e-007 6 > 200 (*) 1 2e-007 7 > 200 (*) 1 le-007 8 > 200 (*) 1 6e-007 9 > 200 (*) 1 4e-007 10 > 200 (*) 1 3e-007 11 > 200 (*) 1 3e-007 12 > 200 (*) 1 le-006 13 > 200 (*) 1 2e-006 14 > 200 (*) 1 2e-006 15 > 200 (*) 1 2e-006 16 > 200 (*) 1 le-006 17 > 200 (*) 1 le-006 18 > 200 (*) 1 2e-006 19 > 200 (*) 1 2e-006 20 > 200 (*) 1 le-006 21 > 200 (*) 1 le-006 22 > 200 (*) 1 le-006 23 < 200 (*) 2 4e-007 24 < 200 (*) 2 5e-007 25 < 200 (*) 2 3e-007 26 < 200 (*) 2 3e-007 27 < 200 (*) 2 3e-007 28 < 200 (*) 2 4e-007 29 < 200 (*) 2 2e-006 30 < 200 (*) 2 le-006 1) no verification of the calibration factors via test predictions of spectral data was performed in this calibration 2) statistical parameter generated by the software 3) rank of the (descriptive) calibration factor 4) residual = calculated value indicating the deviation of a spectral prediction based on calculated calibration factors 5) component of the principal component analysis algorithm of the software 6) radius generated in a plot of calibration factors 2 against themselves (*) ppb PCB per lOOg fat As a result of this Cluster analysis unknown samples can be screened whether they contain PCB's above 200 ppb or not.

Example 2: Partial least squares analysis SESAME software, version 3.1 Beta 1 commercialised by Bran+Luebbe GmbH was used in this example. On the measured spectral data a conversion from reflectance data to absorbance data has been applied and PLS (partial least squares regression was used as calibration calculation as a quantitative analysis. 26 spectra have been selected for calculating calibration factors and predicting the 8 non- selected spectra. According to the cross validation function of the software these calculations were made for all combinations of the spectra and the best set of 26 calibration spectra were selected and a selection of the corresponding calibration factors was used for predictions for all of the 32 recorded spectra. With a selection of 4 calibration factors a regression correlation coefficient of 0. and a standard error of estimate of 212 ppb PCB per 100g fat was determined. However the highest deviation within the predictions for the 32 spectra used in this experiment was only 184 ppb PCP per 100g fat. The data are given in table 3.

Table 3: Partial least squares analysis No of spectra: 32 No. of data points: 701 Wavelength range 1100.. 2500nm, 701,2 nm steps No of properties: 1 CALIBRATION: Calibration method: PLSR Calibration set: 26 spectra 1-12 18-21 ' Selected property: PCB Property range: 0 to (*) Selected Wavelength 1100.. 2500nm, 701,2 nm steps range Transformation: Absorbance -Absorbance

Table 3 : Partial least squares analysis (continued) REGRESSION RESULTS CROSS VALIDATION Selected CROSS VALIDATION SPECTRA: 26 Number of factors : up to 8 Number of factors SEPcv2) (*) 0 1002.04 1 739.447086 2 340. 903536 3 256.

4 237.891447 5 253.757295 6 272.210877 7 301.

8 292.670430 OUTLIER DETECTION: Number Predicted Actual Difference T'H'D'S' 191, 341, 42 (*) 184 (*) 157.42 (*) 0.1.0.3.

Multiple correlation coefficient: 0.980227 Standard error of estimate: 212. (*) Standard error of cross variation: (*) Index of systematic variation7' : 217958 Index of random variation7) : 17148. 2 1) numbers indicating the spectra according corresponding numbers in table 1 2) calculated standard error of prediction in ppb PCB per 100g fat 3) T : student's t test carried out for each spectrum representing the residual error, i. e. how closely the reference value matches the predicted value.

4) H : measure of how strong a particular spectrum influences the resulting regression model corresponding to the multidimensional distance of a spectrum to the regression line 5) D: Cook's D, representing properties of T and H parameters 6) S : spectral reconstruction error, which is obtained by trying to recalculate the original spectrum from the selected factors. (A spectrum is suspected to be an S outlier if the 'S'value is greater than a pre-defined limit).

7) statistical parameters, generated by the software representing the variation in the calibration samples/measurements (*) ppb PCB per 100g-fat

Example 3: Multiple regression SESAME software, version 3.1 commercialised by Bran+Luebbe GmbH was used in this example. On the measured spectral data a conversion from reflectance data to absorbance data has been applied and MLR (multiple linear regression) was used as calibration calculation as a quantitative analysis.

The same of cross validation as in example 2 was used to select calibration spectra. Regression correlation coefficients were calculated for MLR-predictions of the 32 spectra using all possible combinations of sets of three different wavelengths and corresponding correlation factors. The best Regression correlation coefficient determined was 0.9888501 with a standard error estimate of 158 ppb PCB per 100g fat. However the highest deviations within the predictions for the 32 spectra used in this experiment was only 49 and 50 ppb PCB per 100g fat. The data are given in table 3.

Table 3: Multiple regression No of spectra: 32 No of data points : 701 Wavelength range 1100.. 2500nm, 701,2 nm steps No of properties: 1 Properties: PCB Calibration method: MLR-Combination Search Calibration set: 26 spectra 1-13 18-21 24-32 Selected property: PCB Property range: 0 to (*) wavelength range 1100.. 2500nm, 701,2 nm steps Transformation: Absorbance -Absorbance REGRESSION RESULTS-search for combinations of 3 wavelengths best 5 wavelength 1st 2nd 3rd 4th 5th combinations [nm]:----------------------------------- 1730 1730 1730 1376 1376 1740 1740 1740 1384 1384 2260 2264 2258 2184 2180 Regression 0.988501 0.988177 0.988003 0.987917 0.987916 correlation coefficient: Standard Correlation Mean3) Deviation factor Wavelengths : 1730nm 0.404981 0.006661-0.597280 1740nm 0.331089 0.003181-0.350817 2260nm 0.0.006141 0.841206 Property: PCB4 1052. 88 (*) 982. 58 (*) Regression tu) value of Wavelength coefficient coefficient Intercept59-40269 1730nm-748676-12. 786947 1740nm 1.23e+006 14. 179732 2260nm-125902-5. 005744 Degree of Sum of Mean F Source of variation''freedom squares squares Value Attributable to 38) 2. 36e+007 7. 864+006 313.365 regression Deviation from 22 551926 25087.5 regression Total 25 2.

Table 3: Multiple regression (continued) OUTLIER DETECTION: Num-Predicted Actual Difference'T''H''D' ber') 30 2744. 2 (*) 2695 (*) 49. 19205 (*) 0.3.29705 0.

31 2644.9 (*) 2695 (*)-50. 09678 (*)-0.39619 3.14320 0.12282 Multiple correlation 0.988501 coefficient: Standard error of 158.39 (*) estimate : Index of systematic 351336 variation13) : Index of random 1.4e+006 variation13) : E (H)) : 0.115385 1) numbers indicating the spectra according corresponding numbers in table 1 2) calculated standard error of prediction in ppb PCB per 100g fat 3) Lambert-Beer g'coefficients 4) Mean value and deviation (variance) in the calibration measurements 5) offset of the regression line in the multi-linear regression equation 6) Student's t 7) Parameter set representing the variance in the calibration measurements 8) Corresponding to the number of wavelength selected for calibration 9) Representative for scattering 10) T: student's t test carried out for each spectrum representing the residual error, i. e. how closely the reference value matches the predicted value.

11) H : measure of how strong a particular spectrum influences the resulting regression model corresponding to the multidimensional distance of a spectrum to the regression line 12) D: Cook's D, representing properties of T and H parameters 13) statistical parameters, generated by the software representing the variation in the calibration samples/measurements 14) number of wavelength selected for calibration : number of measured calibration spectra (*) ppb PCB per 100g fat

The present invention enables to predict the PCB content in pork fat samples via NIR measurement and can also be used to screen pork fat samples for the PCB level e. g. if they contain above 200 ppb or less than 200 ppb PCB per 100 gram fat in a few seconds, which is a major improvement compared to prior art and to gas chromatography analysis.