3.4. Model fitting and ANOVA
The data obtained for color removal were fitted to various mod- els including linear, two factorial, quadratic and cubic. High and low levels of each factor were coded as 1 and −1, respectively, and the mean value was coded as zero. The analysis of variance shows that the process is most suitably described with a cubic polynomial model. The equation of the model in terms of coded factors is as follows:
Color removal (%) = +11.32−47.47 A−6.28B−17.73 C + 4.10B2
+15.83 C2+55.59 A3+6.16B3 (3)
where A is the coagulant dose, B the pH, and C the dye concentra- tion.
The ANOVA for the model is presented in Table 2. By using an F-test, the significance and lack of fit of the model were evaluated. According to ANOVA, the lack of fit was not significant at the 95% confidence level. The F-value of the model (13.91) with a P-value less than 0.0001 implied that the model was significant. The small P- value and a high coefficient of determination (R2 = 0.8903) showed the suitability of the model for representing the real relationship among the variables.
The significance of each parameter in the model was examined by testing the null hypothesis. If the P-value is less than the signif- icance level, the null hypothesis is rejected, which means that the term is significant (Lee et al., 2010). Some of the insignificant terms of the model were not eliminated to maintain the model hierarchy and obtain a not significant lack of fit. Eliminating all insignificant terms causes that the adjusted R2 (a measure of the amount of variation around the mean explained by the model, adjusted for the number of terms in the model) decreases and the lack of fit
becomes significant (Chaibakhsh et al., 2010). The results show that the coagulant dose and dye concentration significantly affect the color removal performance of O. basilicum and the dye con- centration is the most significant parameter. However, pH (within the selected range) is not a significant factor in the color removal efficiency of the natural coagulant. The parameters also have no interactive effects on the response. This means that the modifica- tion of one of the factors does not affect the other factors inside the working region. Negative coefficient values of the individual factors, A, B and C, indicate that the factors negatively affect the response and by increasing these terms the color removal percent- age may decrease.
Fig. 6 shows the experimental versus predicted color removal percentage obtained from the Eq. (1). The linear distribution is indicative of a well-fitting model. The normal probability plot is also presented in Fig. 6. The plot shows that the residuals follow a nor- mal distribution. The generated mathematical model can be used for prediction of the color removal percentage within the given range of the main parameters.
It should be mentioned that the data obtained for COD reduction could not be fitted well to the models including linear, two factorial, quadratic and cubic. It seems that using a suitable experimental design or employing other nonlinear modeling techniques such as
neural networks, fuzzy logic or a combination of these approaches could be useful for this aim (Sanayei et al., 2014).