Limits of Detection and Quantification in Comprehensive

Jul 1, 2012 - Showing 1/2: ac3010204_si_001.pdf ... his help in developing user-friendly GUIs for various MATLAB scripts used in this and ..... [ACS F...
0 downloads 0 Views 3MB Size
Download PDF
Article pubs.acs.org/ac

Limits of Detection and Quantification in Comprehensive Multidimensional Separations. 1. A Theoretical Look A. Paulina de la Mata† and James J. Harynuk*,† †

Department of Chemistry, University of Alberta, Edmonton, Alberta, T6G2G2 Canada S Supporting Information *

ABSTRACT: Comprehensive multidimensional separations (e.g., GC×GC, LC×LC, etc.) are increasingly popular tools for the analysis of complex samples, due to their many advantages, such as vastly increased peak capacity, and improvements in sensitivity. The most well-established of these techniques, GC×GC, has revolutionized analytical separations in fields as diverse as petroleum, environmental research, food and flavors, and metabolic profiling. Using multidimensional approaches, analytes can be quantified at levels substantially lower than those possible by one-dimensional techniques. However, it has also been shown that the modulation process introduces a new source of error to the measurement. In this work, we present the results of a study into the limits of quantification and detection (LOQ and LOD) in comprehensive multidimensional separations using GC×GC and the more popular “two-step” integration algorithm as an example. Simulation of chromatographic data permits precise control of relevant parameters of peak geometry and modulation phase. Results are expressed in terms of the dimensionless parameter of signal-to-noise ratio of the base peak (S/NBP) making them transportable to any result where quantification is performed using a two-step algorithm. Based on these results, the LOD is found to depend upon the modulation ratio used for the experiment and vary between a S/NBP of 10−17, while the LOQ depends on both the modulation ratio and the phase of the modulation for the peak and ranges from a S/NBP of 10 to 50, depending on the circumstances.

P

evaluated at a 95% confidence interval, and the value chosen for k is 3 or 10, respectively.2 Nowadays the application of comprehensive multidimensional separations is increasing, due to the many advantages that these techniques offer, such as vastly increased peak capacity and separation power, and improvements in sensitivity. The most well-established of these techniques is comprehensive two-dimensional gas chromatography (GC×GC); it has revolutionized analytical separations in fields as diverse as petroleum, environmental research, food and flavors, and metabolic profiling. GC×GC was invented in the early 1990s,5 and in recent years several instruments have become commercially available. The exact workings of comprehensive multidimensional separations are beyond the scope of this work and are discussed elsewhere.6−8 Comprehensive multidimensional techniques provide excellent quantitative and qualitative data, and it has been demonstrated many times that using these techniques, analytes can be identified and quantified at levels substantially lower than those possible using one-dimensional techniques.9 However, it has also been shown that the modulation process at the heart of these techniques introduces a new source of error.10 During modulation, a single analyte peak eluting from

rimary goals of the analytical chemist are ascertaining the presence and quantifying the abundance of a compound in a matrix. The reason for the analysis may be product quality control or trace analysis in a regulatory scenario in a field such as the environment, foods, pharmaceuticals, or personal care products, to mention a few. In many of these cases, the analyst is required to push the limits of the instrumentation to quantify or at least confidently detect analytes at the lowest possible concentrations. Thus, in trace analytical situations, the performance characteristics of Limit of Detection (LOD) and Limit of Quantification (LOQ)1 are of paramount concern. These terms have been studied in great detail for numerous techniques and defined in several ways; however, the most common definitions are those proposed by the International Union of Pure and Applied Chemistry (IUPAC)2 and the International Organization for Standardization (ISO).3 LOD is defined by IUPAC as “the minimum single result which, with a stated probability, can be distinguished f rom a suitable blank value”. And LOQ as “the minimum quantif iable value”.4 Both limits can be calculated using eq 1 yL = ybaseline + k × sbaseline

(1)

where yL is the measured signal, ybaseline is the mean of the blank measurements, sbaseline is the standard deviation of the blank measurements, and k is a factor chosen according to the confidence level desired. Usually, the LOD and LOQ are © 2012 American Chemical Society

Received: April 17, 2012 Accepted: June 30, 2012 Published: July 1, 2012 6646

dx.doi.org/10.1021/ac3010204 | Anal. Chem. 2012, 84, 6646−6653

Analytical Chemistry

Article

Figure 1. (A) First-dimension Gaussian peak simulated with MATLAB. (B) Second-dimension peaks using a modulation period of 3 and a phase of 0°, after processing with ChromaTOF software.

calculated LOD for sterols as the concentration that would give a peak equivalent to 3 times the baseline noise.16 Herein, we probe fundamental limits of quantitation in GC×GC, using simulated chromatographic data to investigate the roles of factors such as MR, ϕ, 1w, 2w, and 1δtr. These simulated data are then processed using common commercial software which relies on a two-step integration approach.

the primary dimension is fractionated into a series of seconddimension peaks, resulting in a base peak (i.e., the tallest modulated peak or slice) and multiple subpeaks. The extent of the error will depend on several factors including the modulation ratio (MR), the abundance of the analyte, the modulation phase (ϕ), the width of the first-dimension peak (1w), and the magnitude of the shifts observed in the firstdimension retention time (1δtr). MR is defined as the ratio of the 1 w of the peak at the base to the modulation period (PM),11 while ϕ is defined as the position of the peak apex in relationship to the modulation pulse sequence. For a symmetrical 1D peak, if the subpeak areas have a symmetrical distribution around a single, maximal base peak, then the modulation is in-phase (ϕ = 0°). Conversely, if the subpeak areas have a symmetrical distribution around two base peaks of equal intensity, then the modulation is completely out-of-phase (ϕ = 180°).12 It has been shown that these new variables will collectively influence the fraction of the analyte signal that is actually observed in a given number of subpeaks and thus call into question the accuracy of analytical results when these techniques are used for trace determinations and only some subpeaks are visible.10 Consequently, a precise definition of LOD and LOQ in modulated multidimensional separations is as yet elusive. Exploring the LOD and LOQ in the context of these types of experiments is the goal of the present work. Due to the lack of definitions of LOD and LOQ for modulated multidimensional techniques, a number of approaches or additional criteria have been proposed in the literature. One approach for considering GC×GC-time-of-flight mass spectrometric (TOFMS) data was proposed by Mitrevski et al. They estimated their LOD as a level where the MS search library correctly identified a compound with a minimum match statistic of 800/999.13 Sets of standards have also been used as LOD indicators.14 With GC×GC-nitrogen chemiluminescent detection (NCD), LOD and LOQ have been calculated applying formulas based on the standard deviations at low concentrations: LOD = s·t (Student t-value for replicates) and LOQ = 10·s.15,15b Beens et al.9 demonstrated that the signal intensity of the second dimension has an enhancement of factor of 15 regarding the signal intensity of the first dimension. The authors found that LOD for GC×GC (0.0005%) was lower than GC (0.009%). Using the general definition, Truong et al.

■

EXPERIMENTAL SECTION Simulated GC×GC Data. Simulated GC×GC chromatograms were generated by a laboratory-written script using MATLAB (v.7.120.635 Mathworks, Natick, USA). The model generated a first-dimension (1D) Gaussian peak of specified 1w (12 s for this study) and with a suitable 1tr to obtain the desired ϕ with the specified modulation period (PM). The modulation of the 1D peak was then simulated to generate a series of second-dimension (2D) Gaussian peaks of specified 2w (here 0.2 s). Gaussian-distributed white noise with a standard deviation typical of that observed in GC×GC-FID experiments was added to the simulated signal, which was then saved in a format that could be imported into ChromaTOF (v.4.33; Leco Instruments, St. Joseph, MI) for data processing. Standard integration settings to capture as many trace peaks as possible were used as follows: The baseline offset was set through the middle of the noise (0.5), data were auto smoothed with an expected 1w of 12 s and an expected 2w of 0.2 s. The signal-to noise ratio (S/N) threshold for both the base peak and inclusion of subpeaks was set at 3. These parameters for integration were chosen to offer the best chances of detecting minor peaks, with minimal integration error, as recommended previously.17 For the evaluation of the LOQ, twelve different 1D peak heights were used, spanning from peaks that were not detectible by the software to peaks which were easily quantified. Each combination of MR (values of 1.5, 2, 3, 4, and 6), ϕ (values of 0, 90, and 180°), and height was simulated ten times. To account for the effect of shifting peaks in 1D, a small 1δtr was applied to the 1D peak prior to modulation. The magnitude of the shift was calculated as a random number chosen from normally distributed values having μ = 0 and σ = 0.6 s. The standard deviation of 600 ms was chosen as this represented the typical value observed for moderately retained peaks across a series of typical GC and GC/MS chromatograms. 6647

dx.doi.org/10.1021/ac3010204 | Anal. Chem. 2012, 84, 6646−6653

Analytical Chemistry

Article

Figure 2. Simulated GC×GC peaks, integrated with Leco software: (A) peak with S/NBP of 50 and (B) peak with S/NBP of 9.

Table 1. Integration Results (Total Peak Area) for 10 Replicate Simulations, MR = 4, ϕ = 0° av S/NBPa 216.8 161.6 106.9 85.9 53.5 43.8 32.8 21.6 17.8 13.2 11.2 9.3 a

detected peak area 38999 29018 17163 15985 9798 6124 4683 3223 2566 1818 0 1310

38528 29044 19314 15206 7691 7014 4833 3537 2554 2340 1575 0

38526 29407 19590 15392 7828 6697 4864 3347 2547 1624 1692 1263

38197 29367 18022 15660 8118 6100 4838 3325 2611 1906 1455 1208

38222 28306 19747 15765 8010 6911 4857 3347 2795 2198 1588 1257

38554 28762 19731 15677 8841 7655 4979 3345 2684 2025 1619 0

38086 29749 19582 15060 9084 6127 4882 3336 2974 2230 1802 1302

38510 29400 19582 15982 8355 6295 4444 3036 2382 2135 1742 1509

38367 28791 19161 15251 8617 6377 4980 2872 3086 1796 1527 0

37541 29021 19473 15609 9829 6375 4971 3142 2934 1842 1551 0

av detected peak area

σ detected peak area

total actual peak area

38353 29087 19137 15559 8617 6568 4833 3251 2713 1991 1455 785

382 412 857 321 767 503 163 190 226 230 522 680

37592 28225 18795 15046 9402 7513 5634 3759 3015 2250 1879 1499

S/NBP: signal-to noise ratio of the modulated base peak.

For LOD estimation, four peak heights were simulated twenty times for each of five MR values (1.5, 2, 3, 4, and 6) at ϕ = 180°. No 1δtr was applied in this case.

(individual subpeaks are integrated and then combined and their areas summed). The parameters relating to this smoothing are critical for the integration of the signal.17 In the interests of space, comparisons of the current results with other integration approaches (e.g., bilinear approaches such as PARAFAC18 or the “watershed” algorithm19) will be presented in subsequent parts of this research, as will comparisons with multidimensional simulated mass spectral data and validation with experimental data. Simulated data were used for this study, as simulation is the only way in which the numerous sources of error may be controlled to study the LOD and LOQ. Figure 1 depicts a simulated 1D peak and the modulated peak profile after addition of noise and data processing. Simulated data were

■

RESULTS AND DISCUSSION The fundamental questions that we are probing involve the evaluation of the limits of detection and quantification in modulated multidimensional experiments. The examples that we are using are GC×GC-FID separations and quantification using ChromaTOF software, which is presently one of the most commonly used commercial packages for GC×GC data interpretation. ChromaTOF incorporates a smoothing step in its data processing and uses a two-step integration approach 6648

dx.doi.org/10.1021/ac3010204 | Anal. Chem. 2012, 84, 6646−6653

Analytical Chemistry

Article

Figure 3. Relative residual of simulated second dimension-peaks: (A) MR of 4 and phase of 180° and (B) MR of 2 and phase of 90°. Error bars indicate one standard deviation.

more important as these shifts cause significant fluctuations in the distribution of peak area among the subpeaks.10,11 In order to estimate values for the LOQ, two approaches were used. First, calibrations were evaluated though considering the residuals for the measured peak areas at each calibration level. Calibration curves were constructed for each combination of MR and ϕ and examined to check for severe outliers (which were removed when they rarely occurred) and verify linear behavior of the data. Then, using only the four highest levels of S/NBP (e.g., 53.8−133.2 for the data shown in Table 1), a calibration was constructed as these peaks had sufficient observable subpeaks and high enough intensities that they were reasonably immune to errors derived from 1δtr or from noise. Thus these points were sure to be in good agreement and exhibit linear behavior. This curve was extrapolated back to the next lower calibration level (e.g., 34.1), and residuals for this lower point were calculated. This process was repeated in an iterative manner, successively adding levels and recalculating the calibration for the calculation of the residuals for the nextlower point. This approach was chosen so that residuals for calibration levels which were in agreement with the calibration curve were not unduly affected by points at lower concentrations which did not agree with the curve. In this interpretation, a bias in the results will be shown by the mean of the residuals drifting from zero, and a lack of precision indicated by an increased spread in the residuals (Figure 3). From the data in Figure 3A, it would be estimated that for this combination of M R and ϕ, the LOQ would be the concentration of analyte providing a S/NBP of about 11, and for Figure 3B the LOQ would be the concentration giving a S/ NBP of around 40. Second, calibration curves were again constructed with the four highest calibration levels, and the curve was extrapolated toward lower levels. Confidence bands for predicted y-values were calculated based on eq 220

evaluated after processing using two parameters: total reported peak area (the sum of the area of the base peak and all included subpeaks) and the S/N of the modulated base peak (S/NBP). The dimensionless parameter of S/NBP was chosen as it will make our results applicable to any situation where the quantification algorithm relies on two-step integration. Furthermore, the use of S/NBP as a metric for evaluating LOD and LOQ has the advantage that it does not depend on 2 w or any other parameter that would require additional experimentation. Finally, as we use S/NBP postprocessing, the results should be equally applicable, whether or not the integration approach incorporates initial noise-reducing strategies (e.g., Gaussian window smoothing) or not. Figure 2 depicts two simulated peaks with different S/NBP after processing. In this case, a portion of the smaller peak’s area was lost because some subpeaks were not detected. Example integration data for 10 replicate simulations of the calibration data for ϕ = 0° and MR = 4 are shown in Table 1. The first column in the table provides the average S/NBP for each calibration level, and the subsequent columns are the total measured peak areas for each replicate simulation. The last three columns present the average detected peak area for each calibration level, the standard deviation of detected peak area, and the “actual” total peak area, as determined by integrating a modulated signal with no noise and no 1δtr. When no subpeaks were detected by the software, a zero is indicated for the total peak area. Integration results for all simulations are provided in the Supporting Information. Theoretical Treatment of the Limit of Quantification. The limit of quantification for a technique is the level at which statistical confidence in the results is lost, in other words, below this limit the calibration curve and the accuracy of the measurement are no longer valid. With modulated multidimensional separations, such as GC×GC, the situation is complicated by the modulation process with two new factors which can influence the LOQ: the MR and ϕ of the 1D peak. It has been shown that if at high concentrations an analyte’s response is measured as the sum of 4 subpeaks accounting for >95% of the area, at low concentrations, when only one subpeak is visible, the analytical signal may be based on only 40−60% of the analyte signal.10,11 In effect, this changes the slope of the calibration curve from that expected based on more concentrated standards. Additionally, as the number of observable peaks is lessened, the influence of 1δtr becomes

y0 ̂ = x0b ±

F(α ; 1, n − p)SR2(x0(XT X )−1x0T )

(2)

Here, ŷ0 is the predicted y-value, x0 is the known x-value, and b is the slope of the calibration curve. F(α;1,n-p) is the Fstatistic for calculation of the confidence interval. Here α of 0.9 (90% confidence interval) was used; n is the number of measurements included in the calibration curve, and p is the number of parameters (two). S2R is the variance residual, and X is the matrix of the independent variables. 6649

dx.doi.org/10.1021/ac3010204 | Anal. Chem. 2012, 84, 6646−6653

Analytical Chemistry

Article

Figure 4. Confidence limits using a phase of 0°: (A) MR of 1.5 and (B) MR of 4. Blue points are included in the calibration; red points are deemed to fall below the LOQ.

Figure 5. Confidence limits using MR = 3: A, B, and C: ϕ = 0, 90, and 180°, respectively. Blue points are included in the calibration; red points are deemed to fall below the LOQ.

would be included in the calibration curve. Figure 4 presents the effect of MR at ϕ = 0°. When MR = 1.5 (Figure 4A) the calibration fails at a S/NBP of 10, while for MR = 4 (Figure 4B) the LOQ is at S/NBP of 30. It is apparent that at constant ϕ, the LOQ becomes higher as MR increases, which can be explained by the fact that with more subpeaks, the individual areas will be smaller, and thus the overall detectability of the individual subpeaks will be lessened.

Points were then added iteratively to the calibration curve by checking to see if the highest excluded level fell within the confidence bands for the calibration. If the measured responses fell within the confidence bands, the points were included in the calibration, and the confidence intervals were recalculated to evaluate the next lower level. If the points were falling outside of the confidence bands, this was an indication that the calibration was no longer valid at this level and now lower levels 6650

dx.doi.org/10.1021/ac3010204 | Anal. Chem. 2012, 84, 6646−6653

Analytical Chemistry

Article

Figure 6. Evaluation of LOQ on the basis of detected peak area vs actual peak area using MR = 3: A, B, and C: ϕ = 0, 90, and 180°, respectively. Blue points are included in the calibration; red points are deemed to fall below the LOQ; error bars = ±1σ.

The effect of ϕ at constant MR is illustrated in Figure 5. Here, it is seen that the LOQ is higher when peaks are in-phase (ϕ = 0°) than when they are out of phase. This is explained by the presence of a single maximum peak for in-phase peaks.21 Thus, the intensities of the subpeaks are lower than when the peaks are partially or fully out-of phase, as in Figure 5B and 5C. This has two effects on the data. First, with smaller subpeaks, they drop to unobservable levels more quickly, resulting in a smaller observable fractional area of the peak, and when only one central subpeak is visible, its area is much more affected by 1δtr than when there are two observable central peaks. With two central peaks, the individual areas will shift, but the sum of the two areas will remain relatively constant, and given that the two peaks have similar areas, they will usually either both be quantified or both missed. More traditionally, one would view a limit of quantification as the concentration of an analyte where the correlation between detected area and true concentration fail. Here, we can approximate the “true concentration” with the integrated peak area of a noisless modulated signal with no peak shift (e.g., final column in Table 1). Figure 6 presents plots of detected peak area vs actual peak area for the conditions shown in Figure 5. The calibration slopes were calculated on the basis of the higher concentrations (blue points), and the points deemed to be below the LOQ are plotted in red. Points were deemed to be below the LOQ on the basis of the distance from the predicted slope and/or a marked increase in the error of the measurement as indicated by the error bars (±1σ). Whether the data are evaluated on the basis of Figure 5 or Figure 6, the indicated position of the LOQ will be the same in terms of the simulated calibration level where the calibration starts to fail. Thus, we have elected to continue using the S/NBP as it is a more easily accessible and generally applicable metric.

Calibration curves for all combinations of MR and ϕ investigated are provided in the Supporting Information. Figure 7 depicts the dependence of S/NBP at the LOQ on the MR and

Figure 7. Flat surface plot for LOQ results. Dark blue represents S/NBP = 50 and dark red represents S/NBP = 10.

ϕ of the peak, demonstrating the lack of a uniquely applicable LOQ. For a peak that is in-phase (ϕ = 0°), the LOQ depends on the MR; however, for out-of-phase peaks (ϕ = 180°) the LOQ is independent of MR. Theoretical Treatment of the Limit of Detection. The LOD is the level at which one can be certain that the signal response is due to the presence of the analyte and not background noise. For determination of the LOD, peaks were initially simulated at ϕ = 180°, as this generates a pulse sequence with two maximal peaks, representing the worst-case scenario in terms of the height of subpeaks for a constant 1D 6651

dx.doi.org/10.1021/ac3010204 | Anal. Chem. 2012, 84, 6646−6653

Analytical Chemistry

Article

peak area. In successive simulations, 1D peak areas were decreased and their modulation simulated. Only Gaussian white noise was added without any 1δtr, and each level was simulated 20 times. The percentage of peaks detected by the software is presented in Table 2. For MR = 2−6 a S/NBP of between 9 and 10 is needed to ensure that the peak is detected 95% of the time, while for a MR of 1.5 the S/NBP could be slightly lower.

heavily on the MR and can be regarded as the concentration of analyte providing a peak with a S/NBP ranging from 10 to 17, depending on the MR. The LOQ, on the other hand, depends on both ϕ and MR and is defined as the concentration of analyte that provides a S/NBP of between 10 and 50, depending on the circumstances. These are conservative theoretical limits for data processed using a two-step integration approach, while allowing for modest shifts in primary retention time. In practice it is recommended that chromatographers evaluate the phase of their analyte peaks as well as the stability of the phase over a series of analyses. If the phase is unstable, than the LOQ should be evaluated as the concentration of analyte providing a S/NBP that satisfies the conditions we have set forth for the worst case scenario. Comparing these theoretical results with experimental data will be presented in future work, as will comparisons of this integration approach with other integration approaches (i.e., watershed algorithm and bilinear chemometric approaches). The LOD and LOQ as defined here in terms of S/NBP do not depend heavily on 2w, though as peaks broaden in 2D, the LOD and LOQ will increase due to the increased mass of anlayte required to generate the required S/NBP. It is likely that LOQ and possibly LOD will depend on peak shape in both dimensions, a study which is also reserved for future work.

Table 2. Percentages of Peaks Detected Using ϕ = 180° S/NBPb MRb

95 >95 >95 >95

a

S/NBP 9.5−10.5 is needed for 95% peak detection for MR = 6. bMR: modulation ratio, S/NBP: signal-to noise ratio of the modulated base peak.

In order to determine the LOD for each MR, the 1D peak area required to generate modulated peaks that were detected 95% of the time when out of phase was used to simulate the modulation of the same peak when it was in phase. These peaks were then simulated 20 times, noise added, and the chromatograms processed. The S/NBP measured in this case (Table 3) is thus defined as the LOD for a given MR (or more

■

S Supporting Information *

Figures 1−10 and a table. This material is available free of charge via the Internet at http://pubs.acs.org.

■

Table 3. S/NBP for ϕ = 0° Given a 1D Peak Area That Will Generate Detectable Peaks When ϕ = 180° S/NBP

ASSOCIATED CONTENT

AUTHOR INFORMATION

Corresponding Author

a

MRa

range

average

6 4 3 2 1.5

9.1−10.8 10.7−11.9 13.3−16.0 14.8−17.7 14.9−18.9

10 11.3 14.6 16.2 16.9

*Phone: +1.780.492.8303. Fax: +1.780.492.9231. E-mail: [email protected]. Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS The authors wish to thank the Natural Sciences and Engineering Research Council of Canada (NSERC) and Alberta Innovates Technology Futures for funding this research. A. Wang is gratefully acknowledged for his help in developing user-friendly GUIs for various MATLAB scripts used in this and other research in our group.

a

MR: modulation ratio, S/NBP: signal-to noise ratio of the modulated base peak.

correctly, the concentration of analyte generating a peak with this S/NBP). To rationalize this, consider a peak with MR = 3, ϕ = 0, and S/NBP = 10. This peak will likely be detected by the software; however, if the peak shifts its 1tr, the intensity of the base peak will decrease. Consequently one cannot be confident of the ability to detect the peak this level in every analysis, if one allows for the possibility of 1δtr. However, the same peak at a S/NBP of 14 will almost certainly be detected, even if the peak shifts completely out of phase. The independence of S/NBP from 2w was verified by simulating a series of peaks as for the data in Tables 2 and 3 for the MR = 3 case but using a 2w of 0.1 s. The percentage of peaks detected in each case was identical to when a 2w of 0.2 s was used, and the results were identical to those for MR = 3 in Table 3, with a range of S/NBP of 12.5−16.3 and an average of 14.4.

■

REFERENCES

(1) Danzer, K. Analytical Chemistry, Theoretical and Metrological Fundamentals, 1st ed.; Springer: Germany, 2007. (2) IUPAC, Compendium of Chemical Terminology, 2nd ed.; McNaught, A. D.m Wilkinson, A., Eds.; Blackwell Scientific Publications: Oxford, 1997; p 399. (3) ISO/IEC, Guide 99: International vocabulary of metrology- Basic and general concepts and associated terms (VIM); ISO: Geneva, 2007. (4) IUPAC. Pure Appl. Chem. 1995, 67 (10), 1699−1723. (5) Phillips, J. B.; Xu, J. J. Chromatogr., A 1995, 703 (1−2), 327−334. (6) Górecki, T.; Harynuk, J.; Panić, O. J. Sep. Sci. 2004, 27 (5−6), 359−379. (7) Francois, I.; Sandra, K.; Sandra, P. Anal. Chim. Acta 2009, 641 (1−2), 14−31. (8) Li, Y.; Wojcik, R.; Dovichi, N. J. J. Chromatogr., A 2011, 1218 (15), 2007−2011. (9) Beens, J.; Boelens, H.; Tijssen, R.; Blomberg, J. High Resolut. Chromatogr. 1998, 21 (1), 47−54.

■

CONCLUSIONS We have provided estimates of both the LOQ and LOD in modulated multidimensional separations. The LOD depends 6652

dx.doi.org/10.1021/ac3010204 | Anal. Chem. 2012, 84, 6646−6653

Analytical Chemistry

Article

(10) Harynuk, J. J.; Kwong, A. H.; Marriott, P. J. J. Chromatogr., A 2008, 1200 (1), 17−27. (11) Khummueng, W.; Harynuk, J.; Marriott, P. J. Anal. Chem. 2006, 78 (13), 4578−4587. (12) Schoenmakers, P.; Marriott, P.; Beens, J. LC GC Eur. 2003, 16 (6), 335−339. (13) Mitrevski, B. S.; Brenna, J. T.; Zhang, Y.; Marriott, P. J. J. Chromatogr., A 2008, 1214 (1−2), 134−142. (14) de Vos, J.; Gorst-Allman, P.; Rohwer, E. J. Chromatogr., A 2011, 1218 (21), 3282−3290. (15) (a) Ozel, M. Z.; Hamilton, J. F.; Lewis, A. C. Environ. Sci. Technol. 2011, 45 (4), 1497−1505. (b) Ozel, M. Z.; Gogus, F.; Yagci, S.; Hamilton, J. F.; Lewis, A. C. Food Chem. Toxicol. 2010, 48 (11), 3268−3273. (16) Truong, T. T.; Marriott, P. J.; Porter, N. A. J. AOAC Int. 2001, 84 (2), 323−335. (17) de la Mata, A. P.; Nizio, K. D.; Harynuk, J. J. Chromatogr., A 2012, in press. (18) (a) Harshman, R. A.; Lundy, M. E. Comput. Stat. Data Anal. 1994, 18 (1), 39−72. (b) Bro, R. Chemom. Intell. Lab. Syst. 1997, 38 (2), 149−171. (19) Reichenbach, S. E.; Ni, M.; Kottapalli, V.; Visvanathan, A. Chemom. Intell. Lab. Syst. 2004, 71 (2), 107−120. (20) Otto, M. Chemometrics, 2nd ed.; Wiley-VCH: Weinheim, 2007. (21) Ong, R. C. Y.; Marriott, P. J. J. Chromatogr. Sci. 2002, 40 (5), 276−291.

6653

dx.doi.org/10.1021/ac3010204 | Anal. Chem. 2012, 84, 6646−6653