Random calibration

Often, it is not quite feasible to control the calibration variables at will. When the process under study is complex, e.g. a sewage system, it is impossible to produce realistic samples that are representative of the process and at the same time optimally designed for calibration. Often, one may at best collect representative samples from the population of interest and measure both the dependent properties Y and the predictor variables X. In that case, both Y and X are random, and one may just as well model the concentrations X, given the observed Y. This case of natural calibration (also known as random calibration) is compatible with the linear regression model... 

This is in contrast to a small 2 sd uncertainty band (Figure 5-35) and random calibration residuals (Figure 5.36) when the baseline was removed. 

Understanding the distribution allows us to calculate the expected values of random variables that are normally and independently distributed. In least squares multiple regression, or in calibration work in general, there is a basic assumption that the error in the response variable is random and normally distributed, with a variance that follows a ) distribution. 

The scatter of the points around the calibration line or random errors are of importance since the best-fit line will be used to estimate the concentration of test samples by interpolation. The method used to calculate the random errors in the values for the slope and intercept is now considered. We must first calculate the standard deviation Sy/x, which is given by ... 

Suppose that you need to add a reagent to a flask by several successive transfers using a class A 10-mL pipet. By calibrating the pipet (see Table 4.8), you know that it delivers a volume of 9.992 mL with a standard deviation of 0.006 mL. Since the pipet is calibrated, we can use the standard deviation as a measure of uncertainty. This uncertainty tells us that when we use the pipet to repetitively deliver 10 mL of solution, the volumes actually delivered are randomly scattered around the mean of 9.992 mL. 

A second example is also informative. When samples are obtained from a normally distributed population, their values must be random. If results for several samples show a regular pattern or trend, then the samples cannot be normally distributed. This may reflect the fact that the underlying population is not normally distributed, or it may indicate the presence of a time-dependent determinate error. For example, if we randomly select 20 pennies and find that the mass of each penny exceeds that of the preceding penny, we might suspect that the balance on which the pennies are being weighed is drifting out of calibration. 

An interesting outgrowth of these considerations is the idea that In r versus K or Vj should describe a universal calibration curve in a particular column for random coil polymers. This conclusion is justified by examining Eq. (9.55), in which the product [i ]M is seen to be proportional to (rg ), with r = a(rg 0 ) - This suggests that In rg in the theoretical calibration curve can be replaced by ln[r ]M. The product [r ]M is called the hydrodynamic volume, and Fig. 9.17 shows that the calibration curves for a variety of polymer types merge into a single curve when the product [r ]M, rather than M alone, is used as the basis for the cafibration. 

Random Measurement Error Third, the measurements contain significant random errors. These errors may be due to samphng technique, instrument calibrations, and/or analysis methods. The error-probability-distribution functions are masked by fluctuations in the plant and cost of the measurements. Consequently, it is difficult to know whether, during reconciliation, 5 percent, 10 percent, or even 20 percent adjustments are acceptable to close the constraints. 

Due to its nature, random error cannot be eliminated by calibration. Hence, the only way to deal with it is to assess its probable value and present this measurement inaccuracy with the measurement result. This requires a basic statistical manipulation of the normal distribution, as the random error is normally close to the normal distribution. Figure 12.10 shows a frequency histogram of a repeated measurement and the normal distribution f(x) based on the sample mean and variance. The total area under the curve represents the probability of all possible measured results and thus has the value of unity. 

The confidence limits of a measurement are the limits between which the measurement error is with a probability P. The probability P is the confidence level and a = 1 - P is the risk level related to the confidence limits. The confidence level is chosen according to the application. A normal value in ventilation would be P = 95%, which means that there is a risk of a = 5 /o for the measurement error to be larger than the confidence limits. In applications such as nuclear power plants, where security is of prime importance, the risk level selected should be much lower. The confidence limits contain the random errors plus the re.sidual of the systematic error after calibration, but not the actual systematic errors, which are assumed to have been eliminated. 

FIGURE 22.2 Calibration curve for OTHdC [Eq. (i), with C = 2.698 for random coiis and C = 4.89 for hard spheres]). (Reprinted from J. Chromatogr. Lib. Ser., 56, 99, Copyright 1995, with permission from Eisevier Science.)... 

The random approach involves randomly selecting samples throughout the calibration space. It is important that we use a method of random selection that does not create an underlying correlation among the concentrations of the components. As long as we observe that requirement, we are free to choose any randomness that makes sense. 

We will now construct the concentration matrices for our training sets. Remember, we will simulate a 4-component system for which we have concentration values available for only 3 of the components. A random amount of the 4th component will be present in every sample, but when it comes time to generate the calibrations, we will not utilize any information about the concentration of the 4th component. Nonetheless, we must generate concentration values for the 4th component if we are to synthesize the spectra of the samples. We will simply ignore or discard the 4th component concentration values after we have created the spectra. 

We will create yet another set of validation data containing samples that have an additional component that was not present in any of the calibration samples. This will allow us to observe what happens when we try to use a calibration to predict the concentrations of an unknown that contains an unexpected interferent. We will assemble 8 of these samples into a concentration matrix called C5. The concentration value for each of the components in each sample will be chosen randomly from a uniform distribution of random numbers between 0 and I. Figure 9 contains multivariate plots of the first three components of the validation sets. 

Option (Valid) presents a graph of relative standard deviation (c.o.v.) versus concentration, with the relative residuals superimposed. This gives a clear overview of the performance to be expected from a linear calibration Signal = A + B Concentration, both in terms of (relative) precision and of accuracy, because only a well-behaved analytical method will show most of the residuals to be inside a narrow trumpet -like curve this trumpet is wide at low concentrations and should narrow down to c.o.v. = 5% and rel. CL = 10%, or thereabouts, at medium to high concentrations. Residuals that are not randomly distributed about the horizontal axis point either to the presence of outliers, nonlinearity, or errors in the preparation of standards. 

A simplified analysis of the effect of particle shape or molecular conformation on SEC calibration has led to the prediction that the more open structure of rigid rod shaped solutes gives a relatively flat SEC-MW calibration curve. As the solute conformation becomes more compact (random-coil to solid-sphere), the SEC-MW calibration curve becomes increasingly steep... 

Figure 5. SEC calibration curves random-coil vs. rigid-rod (8) (SEC column set of several pore sizes, N,N-dimethylacetamide solvent at 80°C)...

Z is a coefficient which relates the concentration of the analyte in the unknown sample to the concentration in the calibration standard, where = bc. R is a residual matrix which contains the measurement error. Its rows represent null spectra. However, in the presence of other (interfering) compounds, the residual matrix R is not random, but contains structure. Therefore the rank of R is greater than zero. A PCA of R, after retaining the significant PCs, gives ... 

The ultimate goal of multivariate calibration is the indirect determination of a property of interest (y) by measuring predictor variables (X) only. Therefore, an adequate description of the calibration data is not sufficient the model should be generalizable to future observations. The optimum extent to which this is possible has to be assessed carefully when the calibration model chosen is too simple (underfitting) systematic errors are introduced, when it is too complex (oveifitting) large random errors may result (c/. Section 10.3.4). 

The P-matrix is chosen to fit best, in a least-squares sense, the concentrations in the calibration data. This is called inverse regression, since usually we fit a random variable prone to error (y) by something we know and control exactly x). The least-squares estimate P is given by... 

We chose the number of PCs in the PCR calibration model rather casually. It is, however, one of the most consequential decisions to be made during modelling. One should take great care not to overfit, i.e. using too many PCs. When all PCs are used one can fit exactly all measured X-contents in the calibration set. Perfect as it may look, it is disastrous for future prediction. All random errors in the calibration set and all interfering phenomena have been described exactly for the calibration set and have become part of the predictive model. However, all one needs is a description of the systematic variation in the calibration data, not the... 

Van der Voet [21] advocates the use of a randomization test (cf. Section 12.3) to choose among different models. Under the hypothesis of equivalent prediction performance of two models, A and B, the errors obtained with these two models come from one and the same distribution. It is then allowed to exchange the observed errors, and c,b, for the ith sample that are associated with the two models. In the randomization test this is actually done in half of the cases. For each object i the two residuals are swapped or not, each with a probability 0.5. Thus, for all objects in the calibration set about half will retain the original residuals, for the other half they are exchanged. One now computes the error sum of squares for each of the two sets of residuals, and from that the ratio F = SSE/JSSE. Repeating the process some 100-2(K) times yields a distribution of such F-ratios, which serves as a reference distribution for the actually observed F-ratio. When for instance the observed ratio lies in the extreme higher tail of the simulated distribution one may... 

---