Howard Mark and Jerome Workman Jr.

Transcription

1 Linearity in Calibration: How to Test for Non-linearity Previous methods for linearity testing discussed in this series contain certain shortcomings. In this installment, the authors describe a method they believe is superior to others. Howard Mark and Jerome Workman Jr. In the previous installment of Chemometrics in Spectroscopy (1), we promised we would present a description of what we believe is the best way to test for linearity (or non-linearity, depending upon your point of view). In the first three installments of this column series (1 3) we examined the Durbin Watson (DW) statistic along with other methods of testing for non-linearity. We found that while the Durbin Watson statistic is a step in the right direction, but we also saw that it had shortcomings, including the fact that it could be fooled by data that had the right (or wrong!) characteristics. The method we present here is mathematically sound, more subject to statistical validity testing, based upon well-known mathematical principles, consists of much higher statistical power than DW, and can distinguish different types of non-linearity from one another. This new method also has been described recently in the literature (4). But let us begin by discussing what we want to test. The FDA/ICH guidelines, starting from a univariate perspective, considers the relationship between the actual analyte concentration and what they generically call the test result, a term that is independent of the technology used to ascertain the analyte concentration. This term therefore holds good for every analytical methodology from manual wet chemistry to the latest high-tech instrument. In the end, even the latest Jerome Workman Jr. serves on the Editorial Advisory Board of Spectroscopy and is director of research, technology, and applications development for the Molecular Spectroscopy & Microanalysis division of Thermo Electron Corp. He can be reached by at: jerry.workman@thermo.com. Howard Mark serves on the Editorial Advisory Board of Spectroscopy and runs a consulting service, Mark Electronics (Suffern, NY). He can be reached via at: hlmark@prodigy.net. instrumental methods have to produce a number, representing the final answer for that instrument s quantitative assessment of the concentration, and that is the test result from that instrument. This is a univariate concept to be sure, but the same concept that applies to all other analytical methods. Things may change in the future, but this is currently the way analytical results are reported and evaluated. So the question to be answered is, for any given method of analysis: Is the relationship between the instrument readings (test results) and the actual concentration linear? This method of determining non-linearity can be viewed from a number of different perspectives, and can be considered as coming from several sources. One way to view it is as having a pedigree as a method of numerical analysis (5). Our new method of determining non-linearity (or showing linearity) also is related to our discussion of derivatives, particularly when using the Savitzky Golay method of convolution functions, as we discussed recently (6). This last is not very surprising, once you consider that the Savitzky Golay convolution functions also are (ultimately) derived from considerations of numerical analysis. In some ways it also bears a resemblance to the current method of assessing linearity that the FDA and ICH guidelines recommend, that of fitting a straight line to the data, as assessing the goodness of the fit. As we have shown (2, 3), based upon the work of Anscombe (7), the currently recommended method for assessing linearity is faulty because it cannot distinguish linear from non-linear data, nor can it distinguish between non-linearity and other types of defects in the data. But an extension of that method can. Expanding a Definition In our recent column we proposed a definition of linearity (2). We defined linearity as The property of data comparing test results to actual concentrations, such that a straight line provides as good a fit (using the least-squares criterion) as any other mathematical function. This almost seems to be the same as the FDA/ICH approach, which we have just discredited. But there is a difference. The difference is the question of 26 Spectroscopy 20(9) September 2005

2 Table I. The results of applying the new method of detecting non-linearity to Anscombe s data sets (linear and non-linear). Parameter Coefficient when using t-value when using Coefficient using t-value using only linear term only linear term square term square term Results for non-linear data Constant Linear term Square term SEE R Results for normal data Constant Linear term Square term SEE R fitting other possible functions to the data; the FDA/ICH guidelines only specify trying to fit a straight line to the data. This also is more in line with our own proposed definition of linearity. We can try to fit functions other than a straight line to the data, and if we cannot obtain an improved fit, we can conclude that the data is linear. But it also is possible to fit other functions to a set of data, using least-squared mathematics. In fact, this is what the Savitzky Golay method does. The Savitzky Golay algorithm, however, does a whole bunch of things, and lumps all those things together in a single set of convolution coefficients: it includes smoothing, differentiation, curve-fitting of polynomials of various degrees, leastsquares calculations, does not include interpolation (although it could), and finally combines all those operations into a single set of numbers that you can multiply your measured data by to get the desired final answer directly. For our purposes, though, we don t want to lump all those operations together. Rather, we want to separate them and retain only those operations that are useful for our own purposes. For starters, we discard the smoothing, derivatives, and performing a successive (running) fit over different portions of the data set, and keep only the curvefitting. Texts dealing with numerical analysis tell us what to do and how to do it. Many texts exist dealing with this subject, but we will follow the presentation of Arden (5). Arden points out and discusses in detail many applications of numerical analysis: fitting data, determining derivatives and integrals, interpolation (and extrapolation), solving systems of equations, and solving differential equations. These methods are all based on using a Taylor series to form an approximation to a function describing a set of data. The nature of the data, and the nature of the approximation considered differs from what we are used to thinking about, however. The data is assumed to be univariate (which is why this is of interest to us here) and to follow the form of some mathematical function, although we might not know what the function is. All the applications mentioned therefore are based upon the concept that since a function exists, our task is to estimate the nature of that function, using a Taylor series, and then evaluate the parameters of the function by imposing the condition that our approximating function must pass through all the data points available, because those data points all are described exactly by that function. Using a Taylor series implies that the approximating function that we wind up with will be a polynomial, and perhaps one of very high degree (the degree of a polynomial being the highest power to which the variable is raised in that polynomial). If we have chosen the wrong function, then there might be some error in the estimate of data between the known data points, but at the data points the error must be zero. A good deal of mathematical analysis goes into estimating the error that can occur between the data points. Approximation The concepts of interest to us are contained in Arden s book in a chapter titled Approximation. This chapter takes a slightly different tack than the rest of the discussion, but one that goes exactly in the direction that we want to go. In this chapter, the scenario described above is changed very slightly. There is still the assumption that there is a single (univariate) mathematical system (corresponding to analyte concentration and test reading ), and that there is a functional relationship between the two variables of interest, although again, the nature of the relationship might be unknown. The difference, however, is the recognition that data might have error, and therefore we no longer impose the condition that the function we arrive at must pass through every data point. We replace that criterion with a different criterion the criterion we use is one that will allow us to say that the function we use to describe the data follows the data in some sense. While 28 Spectroscopy 20(9) September 2005

3 other criteria can be used, a common criterion used for this purpose is the least squares principle: to find parameters for any given function that minimize the sum of the squares of the differences between the data and a corresponding point of the function. Similarly, many different types of functions can be used. Arden discusses, for example, the use of Chebyshev polynomials, which are based upon trigonometric functions (sines and cosines). But these polynomials have a major limitation: they require the data to be collected at uniform X-intervals throughout the range of X, and real data will seldom meet that criterion. Therefore, because they also are by far the simplest to deal with, the most widely used approximating functions are simple polynomials; they also are convenient in that they are the direct result of applying Taylor s theorem, whereby Taylor s theorem produces a description of a polynomial that estimates the function being reproduced: Also, as we will see, they lead to a procedure that can be applied to data having any distribution of the X-values. While discussing derivatives, we have noted in a previous column that for certain data a polynomial can provide a better fit to that data than can a straight line (see Figure 6b of reference 8). In fact, we reproduce that Figure 6b here again as Figure 1 in this column, for ease of reference. Higher degree polynomials might provide an even better fit, if the data requires it. Arden points this out, and also points out that, for example, in the non-approximation case (assuming exact functionality), if the underlying function is in fact itself a polynomial of degree n, then no higher degree polynomial is needed in that case, and in fact, it is impossible to fit a higher degree polynomial to the data. Even if an attempt is made to do so, the coefficients of any higher-degree terms will be zero. For functions other than polynomials the best fit might not be clear, but as we will see, that will not affect our efforts here. The mathematics of fitting a polynomial by least squares are relatively straightforward, and we present a derivation here, one that follows Arden, but is rather generic, as we will see: Starting from Equation 1, we want to find coefficients (the a i ) that minimize the sum-squared difference between the data and the function s estimate of that data, given a set of values of X. Therefore we first form the differences: We then square those differences and sum those squares over all the sets of data (corresponding to the samples used to generate the data): The problem now is to find a set of values for the a i that minimizes ΣD 2 with respect to each a i. We do this by the usual procedure of taking the derivative of ΣD 2 with respect to each [1] [2] [3] a i and setting each of those derivatives equal to zero. Note that because there are n + 1 different a i we wind up with n + 1 equations, although we only show the first three of the set: [4a] [4b] [4c] Now we actually take the indicated derivative of each term and separate the summations. Noting that ( i F 2 ) = 2 i F F (where F is the inner summation of the a i X): etc. [5a] [5b] [5c] Dividing both sides of Equations 5 by 2 eliminates the constant term and subtracting the term involving Y from each side of the resulting equations puts the equations in their final form: etc. [6a] [6b] [6c] The values of X and Y are known, since they constitute the data. Therefore Equations 6a c comprise a set of n + 1 equations in n + 1 unknowns, the unknowns being the various values of the a i because the summations, once evaluated, are constants. Therefore, solving Equations 6a c as simultaneous equations for the a i results in the calculation of the coefficients that describe the polynomial (of degree n) that best fits the data. In principle, the relationships described by Equations 6a c could be used directly to construct a function that relates test results to sample concentrations. In practice there are some important considerations that must be taken into account. The major consideration is the possibility of correlation between the various powers of X. We find, for example, that the correlation coefficient of the integers 1 10 with their 30 Spectroscopy 20(9) September 2005

4 squares is a rather high value. Arden describes this mathematically and shows how the determinant of the matrix formed by Equations 6a c become smaller and smaller as the number of terms included in Equation 1 increases, due to correlation between the various powers of X. Arden is concerned with computational issues, and his concern is that the determinant will become so small that operations such as matrix inversion will be come impossible to perform because of truncation error in the computer used. Our concerns are not so severe; as we will see, we are not likely to run into such drastic problems. Nevertheless, correlation effects still are of concern for us, for another reason. Our goal, recall, is to formulate a method of testing linearity in such a way that the results can be justified statistically. Ultimately we will want to perform statistical testing on the coefficients of the fitting function that we use. In fact, we will want to use a t-test to see whether any given coefficient is statistically significant, compared to the standard error of that coefficient. We do not need to solve the general problem, however, just as we do not need to create the general solution implied by Equation 1. In the broadest sense, Equation 1 is the basis for computing the best-fitting function to a given set of data, but that is not our goal. Our goal is only to determine whether the data represents a linear function or not. To this end it suffices only to ascertain whether the data can be fitted better by any polynomial of degree greater than 1, than it can by a straight line (which is a polynomial of degree 1). To this end we need to test a polynomial of any higher degree. While in some cases the use of more terms might be warranted, in the limit we need test only the ability to fit the data using only one term of degree greater than 1. Hence, while in general we might wish to try fitting equations of degrees 2, 3,... m (where m is some upper limit less than n), we can begin by using polynomials of degree 2, that is, quadratic fits. A complication arises. We learn from considerations of multiple regression analysis, that when two (or more) variables are correlated, the standard error of both variables is increased over what would be obtained if equivalent but uncorrelated variables are used. This is discussed by Daniel and Wood (see page 55 in reference 9), who show that the variance of the estimates of coefficients (their standard errors) is increased by a factor of [7] when there is correlation between the variables, where R represents the correlation coefficient between the variables and we use the term VIF, as is sometimes done, to mean variance inflation factor. Thus we would like to use uncorrelated variables. Arden describes a general method for removing the correlation between the various powers of X in a polynomial, based upon the use of orthogonal Chebyshev polynomials, as we briefly mentioned above. But this method is unnecessarily complicated for our current purposes, and in any case has limitations of its own. When applied to actual data, Chebyshev and other types of orthogonal polynomials (Legendre, Jacobi and others) that could be used will be orthogonal only if the data is uniformly, or at least symmetrically, distributed; real data will not always meet that requirement. Because, as we shall see, we do not need to deal with the general case, we can use a simpler method to orthogonalize the variables, based upon Daniel and Wood, who showed how a variable can be transformed so that the square of that variable is uncorrelated with the variable. This is a matter of creating a new variable by simply calculating a quantity Z and subtracting that from each of the original values of X.A symmetric distribution of the data is not required since that is taken into account in the formula. Z is calculated using the expression (see page 121 in reference 9). In Appendix A we present the derivation of this formula: [8] Response Parabola Wavelength Second derivative Figure 1. A quadratic polynomial can provide a better fit to a nonlinear function over a given region than a straight line can; in this case the second derivative of a normal absorbance band. Then the set of values (X Z) 2 will be uncorrelated with X, and estimates of the coefficients will have the minimum possible variance, making them suitable for statistical testing. In appendix A we also present formulas for making the cubes, quartics and, by induction, higher powers of X be orthogonal to the set of values of the variable itself. In his discussion of using these approximating polynomials, Arden presents a computationally efficient method of setting up and solving the pertinent equations. But we are less concerned with abstract concepts of efficiency than we are with achieving our goal of determining linearity. To this end, we point out that Equations September (9) Spectroscopy 31

5 6a c, and indeed the whole derivation of them, is familiar to us, although in a different context. We all are familiar with using a relationship similar to Equation 1; in using spectroscopy to perform quantitative analysis, one of the representations of the equation involved is: which is the form we use commonly to represent the equations needed for performing quantitative spectroscopic analysis using the MLR algorithm. The various X i in Equation 9 represent entirely different variables. Nevertheless, starting from Equation 9, we can derive the set of equations for calculating the MLR calibration coefficients, in exactly the same way we derived Equations 6a c from Equation 1. An example of this derivation is presented in [10]. Because of this parallelism we can set up the equivalencies: a 0 = b 0 a 1 = b 1 X 1 = X a 2 = b 2 X 2 = X 2 a 3 = b 3 X 3 = X 3 etc. and we see that by replacing our usual MLR-oriented variables X 1, X 2, X 3, and so forth with X, X 2, X 3, and so forth, respectively, we can use our common and well-understood mathematical methods (and computer programs) to perform [9] Circle 22 the necessary calculations. Furthermore, along with the values of the coefficients, we can obtain all the usual statistical estimates of variances, standard errors, goodness of fit, etc., that MLR programs produce for us. Of special interest is the fact that MLR programs compute estimates of the standard errors of the coefficients, as described by Draper and Smith (see, for example, page 129 in reference 11). This allows testing the statistical significance of each of the coefficients, which, as we recall, are now the coefficients of the various powers of X that comprise the polynomial we are fitting to the data. This is the basis of our tests for non-linearity. We need not use polynomials of high degree because our goal is not necessarily to fit the data as well as possible. Especially because we expect that well-behaved methods of chemical analysis will produce results that already are close to linearly related to the analyte concentrations, we expect non-linear terms to decrease as the degree of the fitting equation used increases. Thus we need only fit a quadratic, or at most a cubic equation, to our data to test for linearity, although there is nothing to stop us from using equations of higher degree if we choose. Data well-described by a linear equation will produce a set of coefficients with a statistically significant value for the term X 1 (which is X, of course) and non-significant values for the coefficients of X 2 or higher degree. Conclusion This is the basis for our new test of linearity. It has all the advantages we described: it gives an unambiguous determination of whether any non-linearity is affecting the relationship between the test results and analyte concentration. It provides a means of distinguishing between different types of non-linearity, if they are present, because only those that have statistically-significant coefficients are active. It also is more sensitive than any other statistical linearity test including the Durbin Watson statistic. The tables in Draper and Smith for the thresholds of the Durbin Watson statistic only give values for more than 10 samples. As we will see shortly, however, this method of linearity testing is quite satisfactory for much smaller numbers of samples. As an example, we applied these concepts to the Anscombe data (7). Table I shows the results of applying this to both the normal data (Anscombe s X1, Y1 set) and the data showing non-linearity. We also computed the nature of the fit using only a straight-line (linear) fit as was done originally by Anscombe. We also fitted a polynomial using the quadratic term as well. It is interesting to compare results both ways. We find that in all four cases, the coefficient of the linear term is 0.5. In Anscombe s original paper, this is all he did, and obtained the same result, but this was by design the synthetic data he generated was designed and intended to give this result for all the data sets. The fact that we obtained the same coefficient (for X) using the polynomial demonstrates that the quadratic term was indeed uncorrelated to the linear term.

6 The improvement in the fit from the quadratic polynomial applied to the non-linear data indicated that the square term was indeed an important factor in fitting that data. In fact, including the quadratic term gives well-nigh a perfect fit to that data set, limited only by the computer truncation precision. The coefficient obtained for the quadratic term is comparable in magnitude to the one for linear term, as we might expect from the amount of curvature of the line we see in Anscombe s plot (7). The coefficient of the quadratic term for the normal data, on the other hand, is much smaller than for the linear term. As we expected, furthermore, for the normal, linear relationship, the t-value for the quadratic term for the linear data is not statistically significant. This demonstrates our contention that this method of testing linearity is indeed capable of distinguishing the two cases, in a manner that is statistically justifiable. The performance statistics the SEE and the correlation coefficient show that including the square term in the fitting function for Anscombe s non-linear data set, gives, as we noted above, essentially a perfect fit. It is clear that the values of the coefficients obtained are the ones he used to generate the data in the first place. The very large t- values of the coefficients are indicative of the fact that we are near to having only computer round-off error as operative in the difference between the data he provided and the values calculated from the polynomial that included the second-degree term. Thus this new test also provides all the statistical tests that the current FDA/ICH test procedure recommends. It provides information as to whether, and how well, the analytical method gives a good fit of the test results to the actual concentration values. It can distinguish between different types of non-linearities, if necessary, while simultaneously evaluating the overall goodness of the fitting function. As the results from applying it to the Anscombe data show, it is eminently suited to evaluating the linearity characteristics of small data set as well as large ones. Appendix: Derivation and Discussion of the Formula in Equation 8 Starting with a set of data values X i, we want to create a set of other values from these X i such that the squares of those values are uncorrelated to the X i themselves. We do this by subtracting a value Z, from each of the X i and find a suitable value of Z. so that the set of values (X i Z) 2 is uncorrelated with the X i. From the definition of the correlation coefficient, then, this means that the following must hold: [A1] Multiplying both sides of Equation A1 by the denominator of the LHS of Equation A1 results in the much-simplified expression: [A2] We now need to solve this expression for Z. We begin by expanding the square term: We then multiply through: [A3] [A4] Distributing the summations and bringing constants outside the summations: [A4] References 1. H. Mark and J. Workman, Spectroscopy 20(1), (2005). 2. H. Mark and J. Workman, Spectroscopy 20(3), (2005). 3. H. Mark and J. Workman, Spectroscopy 20(4), (2005). 4. H. Mark, J. Pharm. Biomed. Anal. 33, 7 20 (2003). 5. B.W Arden, An Introduction to Digital Computing, 1st ed. (Addison-Wesley Publishing Co., Inc., Reading, MA, 1963). 6. H. Mark and J. Workman, Spectroscopy 18(12), (2003). 7. F.J. Anscombe, Am. Stat. 27, (1973). 8. H. Mark and J. Workman, Spectroscopy 18(9), (2003). 9. C. Daniel and F. Wood, Fitting Equations to Data Computer Analysis of Multifactor Data for Scientists and Engineers, 1st ed. (John Wiley & Sons, New York, 1971). 10. H. Mark, Principles and Practice of Spectroscopic Calibration (John Wiley & Sons, New York, 1991). 11. N. Draper and H. Smith, Applied Regression Analysis, 3rd ed. (John Wiley & Sons, New York, 1998). Since the last term in equation A4 vanishes, leaving: Equation A5 is now readily rearranged to solve for Z: [A5] [A6] 34 Spectroscopy 20(9) September 2005

7 Equation A6 appears to differ from the expression in Daniel and Wood (9), in that the denominator expressions differ. To show that they are equivalent, we start with the denominator term of the expression on page 121 of reference 9: [A7] Similarly, for fourth powers we set up the expression: which gives: [A15] Again, we expand this expression: and separating and collecting terms: Rearranging the last term in the expression: [A8] [A9] [A10] [A16] Again, Equation A16 is cubic in z 4 and can be solved by algebraic methods. For higher powers of the variable we can derive similar expressions. After the sixth power, algebraic methods are no longer available to solve for the z i, but after evaluating the summations, computerized approximation methods can be used. Thus, the contribution of any power of the x-variable to the non-linearity of the data can be tested similarly by these means. And we find that again, the last term in equation A10 vanishes since, leaving: [A11] And upon combining the summations and factoring out X i : [A12] Which is thus seen to be the same as the denominator term we derived in Equation A6: QED By similar means we can derive expressions that will create transformations of other powers of the X-variable, that make the corresponding power uncorrelated to the X variable itself. Thus, analogously to Equation A2, if we wish to find a quantity Z 3 that will make (X i Z 3 ) 3 be uncorrelated with X, we set up the expression: [A13] Which provides the following polynomial in Z 3 : [A14] Equation A14 is quadratic in Z 3, and thus, after evaluating the summations is easily solved through use of the Quadratic Formula. Circle 25