JOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY MARCH 1991 VOL. 6 I09 Use of Partial Least Squares Modelling to Compensate for Spectral Interferences in Electrothermal Atomic Absorption Spectrometry With Continuum Source Background Correction Part 1. Determination of Arsenic in Marine Sediments* Douglas C. Baxter Wolfgang Frech and lngemar Berglund Department of Analytical Chemistry University of UmeA S-90 I 87 UmeA Sweden When arsenic is to be determined in samples containing aluminium such as sediments by electrothermal atomic absorption spectrometry (ETAAS) at the most sensitive line (1 93.7 nm) spectral interferences occur when continuum source background correction is used. In this work the possibilities of resolving the spectral interference from aluminium mathematically using multivariate calibration have been investigated.A calibration set consisting of arsenic and aluminium standard solutions and mixtures of the two is analysed by ETAAS and the absorbance signals obtained are used to construct a partial least squares model. This model is then used to predict the arsenic (and aluminium) concentrations in dissolved sediment samples from their absorbance signal patterns. The multivariate calibration method employed is described and results for the determination of arsenic in two marine sediment reference materials are discussed. Keywords Arsenic determination; electrothermal atomic absorption spectrometry; marine sediment reference material; spectral interference; partial least squares modelling In the determination of arsenic in environmental samples by electrothermal atomic absorption spectrometry (ETAAS) prob- lems may arise in the form of non-spectral and spectral interfer- ence effects. However by utilizing optimal ETAAS conditions including the platform technique,'.? chemical modification2" and evaluation of peak area^,^,^,^ non-spectral interference effects have largely been eliminated.Such conditions are readily accessible and are incorporated in the stabilized tem- perature platform furnace (STPF)2 concept; a collection of ana- lytical conditions which in the majority of instances facilitates interference-free graphite furnace performance. An additional important requirement for the STPF concept is the use of 'Zeeman-effect background correction for all but the simplest situations'.? In general this is the only STPF con- dition that cannot be fulfilled by all ETAAS instruments hence the problem of spectral interferences in the determina- tion of arsenic may arise.By using continuum source background correction Saeed and Thomassens observed spectral interferences from phos- phate matrices in the determination of arsenic. Riley6 noted un- dercorrection errors at the primary arsenic wavelength ( 1 93.7 nm) resulting from aluminium auto-ionization lines (see Table 1 data from reference 7) and proposed using the line at 197.2 nm which is approximately half as sensitive instead. A more recent study by Martinsen er al.x revealed potential problems from concomitant cobalt and iron at the 193.7 cm line and from cobalt and nickel at the alternative 197.2 nm wavelength. (It should be noted that the use of Zeeman-effect background cor- rection eliminates all the aforementioned spectral interference problems.') For environmental samples such as sediments as discussed here spectral interferences from aluminium at the 193.7 nm arsenic line are likely to be the major problem1() as the other species are not present in sufficiently high concentra- tions to cause background correction errors.In this work the possibilities of mathematically resolving the spectral interference from aluminium in the determination of arsenic in sediments at the 193.7 nm line by ETAAS with con- tinuum source background correction are investigated. It is shown that accurate arsenic results for two marine sediment reference materials [ BCSS- 1.Coastal Marine Sediment and MESS- 1 Estuarine Sediment (National Research Council * Presented at the Fifth Biennial National Atomic Spectroscopy Sympo- sium (BNASS). Loughborough. UK 181h-201h July. 1990. Table 1 Aluminium ion vacuum wavelengths in a 1.0 nm interval centred around the arsenic 193.759 nm line' Wavelength/nm Wavelength/nm 193.4503 11" 193.5949 IIIt 193.4713 I1 193.6907 11 193.5840 111 193.9261 I1 193.5863 111 * 11 Singly ionized aluminium. i. 111 Doubly ionized aluminium. Canada)] can be obtained using multivariate calibration based on partial least squares (PLS) m0del1ing~I-l~ of the signal profiles. Theory Problem Formulation and Notation To achieve accurate background correction in ETAAS using a continuum source the non-specific signal component must be a broad band with an unstructured n a t ~ r e .~ The presence of any structure within the spectral bandwidth isolated by the monochromator other than that due to the analyte (and which can be considered negligibleI4) leads to inaccurate back- ground corre~tion,~ i.e. a spectral interference implies that the background corrected absorbance signal contains contributions from both the analyte and some interfering species. Therefore for the determination of arsenic in matrices with high alumin- ium contents such as sediments the problem is one of resolv- ing the contributions of these two species. The notation used in the following discussions is that ma- trices are represented by boldface uppercase italic letters e . g . R and vectors by boldface lowercase italic characters e.g.r. Partial Least Squares Modelling The problem of spectral interferences is frequent in analytical chemistry and various approaches to their resolution have been proposed.15-'x For the present application PLS model- ling1 1-13.17 of the absorbance signals was used. A calibration set was first prepared by ETAAS analysis of standard solutions of arsenic and aluminium and mixtures of the two. The known110 JOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY MARCH 1991 VOL. 6 concentrations in these solutions were arranged in an n x 2 matrix C (n standard solutions and two species arsenic and al- uminium) and the corresponding absorbance signals in an n x rn response matrix R . Each of the n rows of R describes the absorbanc. signal as a function of time and each of the m columns the absorbance at a specific time for the n standard solutions.The goal in PLS is to model R and C as well as possible and to maximize correlation between these matrices thus confer- ring good predictive capabilitie~.~1.12.'7 To achieve this the mean value of each column in R and C (row vectors F and f of dimensions 1 x rn and 1 x 2 respectively) is first subtracted from each variable in the column making the subsequent com- putations-well conditioned. Next the matrices are decomposed R = l F + T P + E (1) C = 1C + UQ + F (2) where P and Q are the loading for the matrices R and C re- spectively and T and U the sc0res.~*-1~ The column vectors 1 contain ones in all positions and have dimensions n x 1. The loadings and scores matrices are of a lower dimension than the R and C blocks and consist of a number of PLS components which describe the systematic variation in R and C.Matrices E and F are residual matrices containing the non-modelled parts of the data. The relation between U and T is given by U = T B + H (3) where B is a diagonal matrix of model regression coefficients12 and H a residual matrix which is minimized in the least- squares sense. Predictions of the concentrations of arsenic and aluminium in an unknown sample are then obtained from the absorbance signal vector (Tunk) inserted into the PLS model in the sequence runk +funk *uunk j C u n k (4) Number of Significant PLS Dimensions The PLS components are calculated one at a time using an iter- ative a l g o r i t h m ~ ~ ~ ~ 2 ~ ~ 7 which extracts the first component to de- scribe the largest amount of variance in the data the second the greatest part of the remaining variance after the first com- ponent and so on.Iteration continues until no significant vari- ance and hence no information remains. The number of PLS components or dimensions required to predict concentrations accurately is an important factor to con- sider in constructing the model. In the ideal situation predic- tion of arsenic and aluminium concentrations should require only two PLS components. However should the two species interact in any way or detector non-linearities be evident then more than two PLS dimensions will be necessary if accurate estimates of concentration are to be made. To establish the optimum number of PLS components the method known as cross-validationlg is used.After each new PLS component is calculated the predictive capabilities are compared with those of the previous lower-dimensioned model. Cross-validation on newly calculated dimensions continues until predictions are no longer improved and the optimum number of PLS compo- nents to use in the model is established. If too few dimensions are included in the PLS model an adequate description of the system under consideration cannot be obtained and biased estimates of concentration will result.I2JO When too many components are used overfitting of the data in the calibration set occurs which generally means that noise in the measurements is given predictive significance. I x20 Experimental Instrumen tation A laboratory-constructed graphite furnace incorporating side- Table 2 Instrumental parameters for the platform equipped ETAAS system. For sample volume see Table 3; modifier volume 5 pl of 0.15% palladium; wavelength 193.7 nm; lamp current* 9W; and spectral band- width 1.0 nm Stage TemperatureTC Time/s Drying I30 40 Atom izationt 2000 6 Clean-out 2400 3 Ashing 900 45 * Westinghouse arsenic electrodeless discharge lamp.Continuum source t Heating rate approximately 2000 "C s-I. Read command selected. background correction using a Varian hydrogen hollow cathode lamp. heated integrated-contact tubes" was installed on the optical bench of a research spectrometer system based on a modified Varian Techtron AA-6 monochromator.21.22 The furnance was heated by a power supply (LL-Elektronik Bygdsiljum Sweden) equipped with an optical feed-back temperature control system.23 Instrumental parameters are given in Table 2.Absorbance data and tube temperature profiles were acquired at 90.9 Hz using an Ericsson personal computer via a Tecmar Labmaster interface. The ETAAS software was obtained from B. Radziuk (Bodenseewerk Perkin-Elmer Germany) and permits storage of background corrected absorbance signals as ASCII files. Reagents and Materials Standard solutions of arsenic and aluminium were prepared in 0.4 mol dm-3 hydrochloric acid from analytical reagents of the highest available purity. All acids employed in the dissolution of sediment samples were of pro analysi quality supplied by Merck and used without further purification. A 0.15% palladi- um (as nitrate) solution was used as chemical modifier throughout being prepared from a Merck ETAAS standard.Spectroscopic-reagent grade argon was utilized as the furnace purge gas. Integrated-contact tubes (I9 x 5.7 mm i.d.) were manufac- tured from single pieces of RWO quality graphite and were coated with pyroltic graphite (Ringsdorff-Werke GmbH Germany). Solid pyrolytic graphite platforms as supplied by Perkin-Elmer were used. Two marine sediment reference materials (BCSS- I and MESS- 1 ) were obtained from the National Research Council Can ad a Dissolution Procedure The sediments were dissolved following the procedure of Stur- geon et ~ 1 . ' ~ After drying to constant mass at 105 O C 0.5 g of sediment was placed in a 100 ml polytetrafluoroethylene beaker and wetted with 4 ml of water.Then 5 ml of concen- trated hydrochloric acid 2 ml of concentrated nitric acid and 5 ml of concentrated hydrofluoric acid were added the beaker was covered and heated for 2 h at about 90 "C on a hot-plate. The cover was removed and the solutions slowly brought to dryness then 5 ml of copcentrated nitric acid 2 ml of concen- trated hydrochloric acid and 100 p1 of concentrated perchloric acid were added to effect dissolution of the organic material. On evaporation to dryness the residue was dissolved in 20 ml of warm 1 mol dm-3 hydrochloric acid and diluted to 50 ml with water. The supernatant liquid was carefully decanted to leave the small amounts of undissolved materiaP behind. Blank dissolutions were performed in parallel.Following dissolution the sediment samples were finally further diluted in 0.4 mol dm-3 hydrocholoric acid giving di- lution factors of 1 k 335 for BCSS- 1 and 1 k 301 for MESS-1. To confirm that no arsenic was lost during the digestion pro- cedure the dissolved sediment samples were also analysedJOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY MARCH 199 I. Table 3 Experimental design for calibration. Four replicates performed at each point marked x in the design; 10 pl blank volumes and 5 pl of standard (plus 5 p1 of blank where appropriate). Manual injections. Note that 5 p1 of the dissolved sediment samples were injected then 5 p1 of the blank or an arsenic standard were added the latter for the results discussed later in Table 5 Arsenic Aluminium concentration/mg I-' concentrat ion/ PE I-' 0 100 200 0 50 100 X X X X X X - - X using a Perkin-Elmer Zeeman 3030 equipped with HGA 600 and AS-60 peripherals.Arsenic was determined using STPF conditions,2 and the results obtained were in agreement with the certified values. Verification that no spectral interferences were observed at the 193.7 nm arsenic line was also made using Zeeman-effect background correction for aluminium concentrations of up to 500 mg 1-I (about three times higher than in the sample digests). No significant arsenic (or alumin- ium) concentrations were observed in the blanks following di- gestion. Data Analysis by PLS A program was written to convert the ASCII files stored by the ETAAS software into a format suitable for evaluation by PLS using the SIMCA 3B pa~kage.2~ An editing program was also used which permitted the averaging of several absorbance signals.'3 The calibration set was constructed by ETAAS analysis of arsenic and aluminium standard solutions according to the ex- perimental design given in Table 3. Background corrected ab- sorbance data were collected from the spectrometer every 11 ms during the atomization step using the ETAAS software. The total number of data points acquired for the calibration set thus amounted to about 1.5 x 104 which cannot be convenient- ly handled by the personal computer therefore two methods of data reduction were used. First only the 240 points in the time interval 0.88-3.52 s were used for calibration purposes as the entire signal appeared in this temporal window (see also Fig.1). Second the four replicates at each calibration point were averaged. The latter procedure had no great effect on the predictive properties as verified by comparing results using PLS models based on both reduced and non-reduced data sets. 0.15 (a) L 0.15 9 0.075 0 2.5 5 VOL. 6 1 1 1 Reduction of the data set by signal averaging did on the other hand decrease the required computation time considerably. The results reported here are all based on the reduced data set as this permits fairly rapid processing (about 2 h in total). For the prediction of concentrations in the sediment samples no signal averaging was used so that an approximate assessment of the variations in the results could be made. A more thor- ough evaluation of the uncertainty in the predicted results in- cluding the effect of variance in the calibration set would require an alternative PLS alg~rithm,'~,'~ and is not considered further here.Results and Discussion Graphite Furnace Conditions Preliminary studies showed that the ashing conditions were not critical as the use of the palladium modifier was efficient in stabilizing arsenic to even higher temperatures than the 900 "C given in Table 2. However at higher ashing tempera- tures the arsenic signal appeared somewhat earlier in time before the tube wall temperature had stabilized during the atomization step. An ashing temperature of 900 "C was thus made to ensure a stablized tube wall temperature for the dura- tion of the absorbance pulses (see Fig. 1). For the atomization temperature 2000 "C was used which is lower than that typically recommended for arsenic deter- mination by ETAAS.2.4 At higher temperatures the arsenic peaks were narrower higher and of smaller area.Aluminium signals were also narrower and higher but were of greater area as expected from the larger degree of ionization at ele- vated temperatures. Furthermore the separation between the arsenic and aluminium peak maxima was reduced and thus the conditions for mathematically resolving these signal components were less favourable at temperatures above 2000 "C. It was also confirmed that none of the other sample compo- nents (phosphates iron cobalt)s.x were present in the dissolved sediments at sufficiently high concentrations to produce spec- tral interferences at the 193.7 nm arsenic line.This was achieved by preparing standard solutions having approximate- ly the same concentrations of these components as are present in the dissolved sediments using the relevant data from the certificates of analysis. The data given in Table 4 indicate that the measurement pre- cision is not very good. This may be partly owing to the use of 0.3 (4 0.3 A I Time/s Fig. 1 Averaged absorbance signals for the calibration set (Tables 3 and 4) (peak-area values are given in parentheses). (u) A 50 pg I-' As (peak area = 0.043); and B 10 mg I-' Al (0.073). ( h ) A 50 pg I-' As plus 100 mg I-' A1 (0.122) and B numerical addition of the two signals in (u) (0.1 16). (c) A 100 pg I-' As (0.074); and B. 200 mg I-' A1 (0.151). ( d ) A 100 pg I-' As plus 200 mg I-' A1 (0.241); and B numerical addition of the two signals in (c) (0.225).All signals have been corrected for the blank. Tube temperature (2000 "C) profile C is also shown in (0)112 JOURNAL OF ANALYTICAL A.TOMIC SPECTROMETRY MARCH 1991 VOL. 6 Table 4 Peak characteristics for the calibration set and sediment samples Peak characteristics* Sample Peak areat Peak height fFak/s 50 pg I-' As 100 pg I-' As 100 mg I-' Al 200 mg 1-' Al 50 pg I-' As + 100 mg I-' Al 100 pgl-IAs+ 200 mg I-' Al BCSS- 1 BCSS-1 + 50 pg 1-' As BCSS-1 + 100 pg I-' As MESS- 1 0.043 f 0.005 0.074 f 0.007 0.073 k 0.006 0.151 f 0.002 0.122 f 0.004 0.241 f 0.004 0. I55 f 0.005 0.194 k 0.017 0.228 f 0.010 0.164 f 0.010 0.072 f 0.005 0.127 f 0.007 0.087 f 0.006 0.161 k0.007 0.140 f 0.005 0.254 f 0.007 0.176 k 0.002 0.273 f 0.0 13 0.388 k 0.0 15 0.179 f 0.014 * Mean value f one standard deviation (n = 4).i Peak areas have been corrected for blanks. I .57 k 0.04 1.51 f 0.01 2.03 f 0.05 2.02 f 0.04 1.77 f 0.01 1.78 f 0.03 1.82 f 0.02 1.64 f 0.07 1.58 f 0.06 1.85 k 0.04 Table 4 the results of the univariate calibration approaches may be rejected.3O Tables 4 and 5 also show data and results for the determina- tion of arsenic in BCSS-1 using univariate calibration by the standard additions method. The peak area result is similar to (or as inaccurate as) that based on the calibration graph proce- dure indicating the lack of non-spectral interference effects in the arsenic determination. This is an important finding for the use of aqueous standard solutions in the multivariate calibra- tion procedure discussed below.Although the use of standard additions and peak height measurements gives an arsenic concentration closer to the certified value (see Table 5 ) the result is still seriously in error. We1z3l has also emphasized that the use of standard ad- ditions cannot correct for spectral interferences. Thus unvari- ate calibration is unsuitable for this application and for the analytical conditions used. The considerable danger of obtain- ing erroneous results by assuming that the peak height is unaf- fected by the presence of a component that causes spectral interference is also evident. manual pipetting but is mostly a result of drift in the arsenic lamp intensity which caused slight base-line shifts leading to some uncertainties in peak evaluation.Univariate Calibration For the instrumental conditions used here the arsenic peak absorbance is observed earlier in time than that of alumin- ium (see Table 4). However the peaks are not sufficiently separated to allow calibration based on the peak height method as has been done in several ETAAS applications where spectral interferences have been Indeed for the results given in Table 4 the arsenic concentrations evaluated on a univariate peak height basis would give results approximately four times higher than the certified values (Table 5 ) . Peak area results are even more seriously overestimated. It can be seen in Table 4 that the times at which the peak ab- sorbances occur (cFak) are significantly earlier for the pure arsenic standards relative to those samples which also contain aluminium.Harnly3O has suggested using various temporal characteristics to assess the accuracy of peak height and area measurements for analytes in the absence and presence of a matrix. On the basis of the differences in rPak apparent in Interaction Effects The experimental design shown in Table 3 allows the inclu- sion of interaction effects between arsenic and aluminium. That such an effect exists can be seen in Fig. 1. Here the average signals for the calibration samples are plotted togeth- er with those obtained by numerically adding the pure compo- nent absorbance profiles. All signals shown have had the average blank signals subtracted. It is obvious that while the leading edges of the signals are similar for the mixed standards (arsenic plus aluminium) and the added absorbance profiles [Fig. l(h) and (41 the peak heights are much greater and the tailing less in the former.This means that the signal shape ob- tained on atomizing arsenic and aluminium together is not simply a linear combination of the individual single-component signals. For this reason it is important for the PLS model to include interaction effects. l7 From the results given in Table 4 and Fig. 1 it is also obvious that there are some non-linearities present in the cali- bration set. Thus the PLS model will require more than two di- mensions to enable accurate predictions of the arsenic concentrations in unknown samples i.e. the system under consideration is not ideal (see Number of Significant PLS Dimensions under Theory).Fortunately the PLS method is capable of handling such non-linearities.I3.l7 Table 5 methods. Data from Table 4 Results for the determination of arsenic in sediments by ETAAS with continuum source background correction using univariate calibration Arsenic concentration/pg g-' Calibration graph Standard additions Sample Dilution factor Peak area Peak height Peak area Peak height Certified value BCSS-1 1 +335 69 46 72 27 1 1 . 1 f 1.4 MESS- 1 1 + 301 66 42 - - 10.6 k 1.2 Table 6 ate calibration* Results for the determination of arsenic and aluminium in sediments by ETAAS with continuum source background correction using multivari- Arsenic concentration@& g-' Aluminium concentration (96) Sample PLS t Certified PLSf Certified BCCS- 1 12.3 f 1.7 11.1 f 1.4 9.77 f 0.10 11.83 f 0.41 MESS- I 9.7 f 1.7 10.6 f 1.2 9.85 k 0.85 I 1.03 k 0.38 * A three-component PLS model was used which explained 99.7% and 99.5% of the variance in the response ( R ) and concentration (C) matrices t Error terms are one standard deviation of the concentrations obtained by fitting four individual absorbance signals to the PLS calibration model.respectively.JOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY. MARCH 1991 VOL. 6 113 Time/s Fig. 2 Averaged absorbance signals for the sediment samples solid line BCSS-1; and broken line MESS-1. Signals have been corrected for the blank 16.7 11.1 13.47 11.83 -in 5.5 1 0 I T 10.6 A 10.19 5 C .- U I c 8 a 12.55 11.03 1 3 5 No. of PLS components Fig. 3 Predicted C arsenic and C aluminium concentrations in ( u ) BCSS- 1 and ( h ) MESS- 1 as a function of the number of PLS components used in the model.Solid line is the certified concentration and broken lines are the 95% confidence limits. Error bars are one standard deviation of the predicted concentrations for four individual absorbance signals fitted to the model Multivariate Calibration by PLS Modelling Average absorbance signals for the sediment samples are shown in Fig. 2 and it can be seen that the shapes closely correspond to those of the mixed arsenic plus aluminium standards in Fig. 1. The concentrations for these two species as determined by a three-component PLS model are reported in Table 6. Fig. 3 shows the predicted arsenic and aluminium concentrations for the two marine sediment reference materials as a function of the number of PLS dimensions employed in the model where it is clear that three components are optimum particularly with respect to arsenic in BCSS- 1 .One of the attractive features of multivariate calibration using PLS modelling is that both the analyte element and the spectrally interfering component can be determined simultane- ously by employing a suitable experimental design for the cali- bration set (see Table 6). However while acceptable results for arsenic are obtained the aluminium concentrations deter- mined are in error. This is probably owing to the presence of residual fluorides in the sample which severely depress the formation of free aluminium atoms32 (and hence ions) under the graphite furnace conditions used.Such an effect cannot be accounted for in a PLS model based on aqueous standard solu- tions as used here. An alternative experimenta1 design making standard additions of both aluminium and arsenic to sediment samples prior to construction of the PLS model,12 might improve the predictive capabilities for aluminium. This would of course complicate the calibration step being considerably more time consuming. Nevertheless the main objective of this work to determine arsenic in sediments accurately by ETAAS with continuum source background correction at the 193.7 nm line where severe spectral interferences from aluminium are present has been realized. Conclusions Multivariate calibration based on PLS modelling can be used to correct for spectral interferences in ETAAS.Although the spectral interference problems observed in the determination of arsenic in sediments can be avoided by selection of the less sensitive 197.2 nm wavelength6.I0 or using Zeeman-effect background correction? such alternatives might not always be convenient or indeed available. Thus the method used here may offer a solution to the general problem of background cor- rection errors in ETAAS. One disadvantage of this method may however lie in the need to know a priori the cause of the spectral interference and design the experiments accord- ingly. Furthermore a computer based data acquisition and storage system and access to suitable software,2s is necessary. This work was supported by the Swedish Centre for Environ- mental Research and the Natural Sciences Research Council.We are indebted to E. Lundberg Norrby Marine Research Laboratory Hornefors Sweden for the provision of the sedi- ment reterence materials and to B. Hiitsch Ringsdorff-Werke Bonn Germany for supplying us with graphite parts. The as- sistance of K. Olsson in dissolving the sediment samples and M. Berglund with the computer programs is also gratefully acknowledged. 1 2 3 4 5 6 7 8 9 10 I I 12 13 14 15 16 References L'vov B. V.. Spectrochirn. Actu. Port B 1978,33 153. Slavin W. Gruphite Fi4rnar.e AAS - A Source Book. Perkin-Elmer Norwalk CT 1984. Ediger R. D. At. Ahsorpt. NeM,sl. 1975 14 127. Schlemmer G. and Welz B. Spectrochirn. Actu. Part B 1986 41 1157. Saeed K. and Thomassen Y. Anal. Chirn. Actu 198 I 130,28 I . Riley K. W.At. Spectinsc. 1982 3 120. CRC Hutidhook of Chemistry utid Physics ed. Weast R . C. CRC Press Boca Raton FL. 62nd edn. 198 1 pp. E206-E2 10. Martinsen I. Radziuk B. and Thomassen Y . . J . A w l . At. Spec,ti-om. 1988.3 1013. Slavin W. and Camrick. G.R.. CRC Crir. Re\.. A w l . Cheni.. 1988 19,95. Bettinelli M. Pastorelli. N. and Baroni. V.. Anul. Chini. Aem 1986. 185 109. Wold S. Ruhe A. Wold. M.. and Dunn. W. J.. 111 SlAM .I. S . i . Stutist. Conrpiit. 1984. 5 735. Geladi. P.. and Kowalski. B. R.. A w l . Ckinr. Ac.tu 1985. 185 I . Baxter. D. C. and Ohman. J.. SpecmxAini. Ac.tu. Purt B. 1990. 45 481. Welz. B.. Atoniic Ahsoiptioii Specrronic>ti.\'. VCH. Weinheim. 2nd edn. 1985. p. 136. Ho C.-N.. Christian G. D. and Davidson. E. R.. A m / . Cheni.. 1978. 50 1 108. Saxberg B. E. H . and Kowalski B. R.. A w l . C/wni.. 1979. 51 1031.114 JOURNAL OF ANALYTICAL ATOMIC SPECTROMETRY MARCH 1991 VOL. 6 17 18 19 20 21 22 23 24 Lindberg W. Person J.-A. and Wold S . Anal. Chem. 1983 55 643. van Veen E. H. and de Lms-Vollebregt M. T. C. Spectrochim. Acra Part B 1990,45,313. Wold S. Technometrics 1978,29,397. Martens H. PhD Thesis Technical University of Norway Trond- heim 1985. Frech W. Baxter D. C. and Hutsch B. Anal. Chem. 1986 58 1973. Lundberg E. and Frech W. Anal. Chem. 1981.53 1437. Lundgren G. Lundmark L. and Johansson G. Anal. Chem. 1974 46 1028. Sturgeon R. E. Desaulniers J. A. H. Berman S. S. and Russel D. S . Anal. Chim. Acra 1982 134,283. 25 26 27 28 29 30 31 32 Wold S. SIMCA-3B Users Manual with Examples Research Group for Chemometrics Department of Organic Chemistry University of Umei 1983. Lorber. A. and Kowalski B. R. J. Chemometrics 1988,2,93. Lorber A. and Kowalski B. R. Appl. Specrrosc. 1988,42 1572. Sampson B. J. Anal. At. Spectrom. 1987,2,447. Jacobson B. E. and Locklitch G. Clin. Chem. 1988,34,709. Harnly J. M. J. Anal. At. Specrrom. 1988,3,43. Welz B. Ft-esenius 2. Anal. Chem. 1986,325,95. Dittrich K. CRC Crir. Rev. Anal. Chem. 1986 16 223. Paper 010292621 Received July 2nd I990 Accepted August 29th I990