We present a plot of the residuals from Table 2. As we will see, the use of the residual plot can be an effective way of comparing the fits of competing models. In fact we will compare the fit of the second-degree model of Eq. In particular, to check the normality assumption of the errors, some computer software packages offer plots of the studentized residuals on normal probability paper.
Still another statistic, the Cp statistic, can be used to compare the fits of competing models. This statistic measures the sum of the squared biases, plus the square residuals the eu2 terms at all data points. The fitting of the second-degree yarn elongation model of Eq. At the conclusion of the example, component 1 was said to blend in a synergistic manner with each of components 2 and 3, while components 2 and 3 were said to blend in an antagnositic way.
These claims were made because of the signs and relative magnitudes of the estimates b12 , b13 , and b The question we ask ourselves now is: If we refit the data using only a first-degree model in x1 , x2 , and x3 , would the first-degree model fit the 15 elongation values as well as the second-degree equation 2. With the 15 yarn elongation values in Table 2. The surface contours generated with the fitted model 2. The sums of squares quantities associated with the fitted first-degree model of Eq.
With Eq. For the second-degree model of Eq.
http://danardono.com.or.id/libraries/2020-05-27/jotaj-how-to-put.php A test that compares the first- and second-degree models in and RPRESS terms of how well they account for or explain the variation in the response values uses the residual sum of squares associated with each of the models. The test statistic is an F-ratio of the form see Section 7. In the F-test of Eq.
Share. View Table of Contents for A Primer on Experiments with Mixtures Book Series:Wiley Series in Probability and Statistics. A Primer on Experiments with Mixtures is an excellent book for one-semester in an easy-to-follow manner that is void of unnecessary formulas and theory.
With the yarn elongation data, the value of the F-ratio in Eq. The decrease in the magnitudes of the individual residuals from fitting the second-degree model is illustrated in Figure 2. Also the differences between the shapes of the two estimated surfaces is seen by comparing the contour plots of Figures 2. Finally, Table 2. Listing the outputs from the fitted linear blending model of Eq. As mentioned later in Section 4.
While both models are acceptable by these rules, our choice is the quadratic model of Eq. Having decided from the F-test in Eq. Such tests determine the significance of the individual parameters, as shown in Appendix 1A and in Section 2. A partial solution to the use of too few component mixtures is provided by the simplex-centroid design, which is introduced in Section 2. Displayed in Table 2. Following the fitted second-degree model are the elongation data values, the predicted elongation values using the second-degree model, the residuals and their leverage values that are discussed later in Section 4.
Constituents: Polyethylene x1 , polystyrene x2 , polypropylene x3. Since the F-test in Eq. A slightly different approach that could have been used in developing an appropriate model form would have been to initially fit the first-degree model and then ask: Does the model adequately fit the observed response values? In this approach we assume initially that the components blend linearly and, upon fitting the firstdegree model, question whether there is evidence that is contrary to this assumption.
If there is evidence that suggests the fitted first-degree model is not adequate, then either additional experiments must be performed that might improve the fit or the form of the model must be changed.
There are several approaches that can be taken to test lack of fit of a fitted model. One approach requires that replicate observations be taken at one or more design points, where the number of distinct design points exceeds the number of terms in the fitted model. When this happens, the residual sum of squares from the analysis of the fitted model can be partitioned into two sums of squares: the sum of squares due to lack of fit of the model and the sum of squares due to pure error, where the latter sum of squares is calculated by using the replicates.
These sums of squares, when divided by their respective degrees of freedom, are then compared in the form of an F-ratio as shown in Eq. When replicate observations are not available, other authors Green, ; Daniel and Wood, ; Shillington, have proposed grouping values of the response that are observed at similar i.
The sweeteners were glycine, saccharin, and an enhancer. Initially six combinations of the three sweeteners were selected for testing and are listed as blends 1 to 6 in Table 2. Each respondent in the population was asked to rate the intensity of sweetness aftertaste using a scale of 1 positively no aftertaste to 30 very extreme aftertaste. As a group, the four check-point locations represent positions that are farthest away distancewise from the original six design points and that when used to test the adequacy of the model would appear to maximize the power of the test.
At each of the four check-point blends, 20 people were asked to score the intensity of aftertaste and the average scores recorded by the additional 80 persons are listed in Table 2. Using the fitted model 2. Fitted to the complete set of 19 data values in Table 2. A test of zero lack of fit of the model 2.
The special cubic model, fitted to the 19 values in Table 2. In other words, the design consists of every nonempty subset of the q components, but only with mixtures in which the components that are present appear in equal proportions. Presented in Figure 2. At the points of the simplex-centroid design, data on the response are collected and a polynomial is fitted that has the same number of terms or parameters to be estimated as there are points in the associated design.
The formulas for the estimates of the first two sets of parameters in Eq. The equality of the formulas in Eqs. Each chemical was applied individually and in combination with each of the others to comprise the four single-component blends, six binary blends, four ternary blends, and the four chemicals together. Each of the 15 chemical treatments was sprayed on three plants in each of four blocks of 45 plants.
Seven days after spraying, the total number of mites on 10 leaves sampled from each plant was recorded. An average was taken across the three plants that received the same treatment and the average value was used as a datum. The average relative percentages averaged across the four replications and the component proportions are presented in Table 2.
The quantities in parentheses below the parameter estimates in Eq. The percentages are used here simply for purposes of illustrating the fitting of model 2. According to the average percentage values in Table 2.
Hypothesis tests are discussed in Sections 2. Plotting estimated response contours for systems with four or more components is not an easy exercise. This is because to represent the surface described by the fitted model in Eq. For example, let us assume that surface contours across the values of the component proportions x1 , x2 , and x3 are desired at three levels 0, 0. The estimated response equation 2. Axial designs, however, are designs consisting mainly of complete mixtures or q-component blends where most of the points are positioned inside the simplex.
To define an axial design, we state the following. This distance is defined in the simplex coordinate system as one unit.
Such a design has been suggested in Cornell A three-component axial design is shown in Figure q 2. With an axial design of the form shown in Figure 2. Given the form of the matrix var b in Eq. This means that when fitting the first-degree model to an axial design of the type considered, the greater the number of components, the more spread the design should be in order to increase the precision of the parameter estimates and reduce the correlations between the estimates.
Here precision refers to the reciprocal of the variance and correlation between pairs of estimates is directly related to the covariance between the pairs. Additional discussion on the use of axial designs is presented in Chapters 4 and 5 when the topic of screening the components is discussed.