ANOVA for Linear Mixture

The ANOVA is where the descriptive statistics and statistical tests are presented. In general, look for low p-values to identify important terms in the model.

Select View > Annotated ANOVA to see the blue annotation text to help interpret the key elements in the ANOVA report.

Right-click in a cell on the report and select Help from the menu for details about that section.

The sections that are available depend on the type of model fit to the data.

This style of ANOVA is used when a linear mixture model is selected. It includes a section for testing component gradients rather than the linear coefficients.

ANOVA Sections

ANOVA: This is the analysis of variance for the linear mixture model. Because the linear coefficients in the Scheffé mixture model are a prediction rather than an effect they are not tested independently. If only the linear terms are present in the model, the individual component effect tests are moved to the Gradients section.

Summary Statistics: The descriptive statistics are used as a secondary check for the usefulness of the model.

Subtract the Predicted R-Squared from the Adjusted R-squared. If the difference is less than 0.2, then the model is fitting the data and can reliably be used to interpolate.

Check the Adequate Precision. If it is greater than 4, then the model has a strong enough signal to be used for optimization.

CV% is used in some industries to judge the capability of a process; lower is better.

Compare the standard deviation to the estimate used when sizing the design (power or FDS).

The mean is the average of the response, and the PRESS is used to calculate other statistics such as the predicted R-squared.

Coefficients: This section provides the confidence intervals around the estimated model coefficients. While most analyses do not require examining these intervals, they can help sort out issues when the analysis doesn’t make sense.

Gradients: This section computes the gradients and component effects. Hypothesis tests for the gradients being different than 0 are shown. Look for low p-values to find the active components.

Equations: As many as three model equations will be presented. One for the pseudo model taken from the coefficient table above, one in the real coding scale, and one in the actual scale. The default analysis computes and tests the pseudo model which is converted to the real model and finally to the actual model.