Model Browser User's Guide

Verifying the Global Model

The next step is to check through the global models to see how well they fit the data.

1. Expand the PS22 local node on the Model Browser tree by clicking the plus sign (+) to the left of the icon. Under this node are four response features of the local model. Each of these is a feature of the local model of the response, which is torque.

   

2. Select the first of the global models, knot.

1. The Response Feature pane appears, showing the fit of the global model to the data for knot. Fitting the local model is the process of finding values for these coefficients or response features. The local models produce a value of knot for each test. These values are the data for the global model for knot. The data for each response feature come from the fit of the local model to each test.

   

3. The response feature knot has one outlier marked. Points with an absolute studentized residual value of more than 3 are automatically suggested as outliers (but included in the model unless you take action). You can use the right-click menu to remove suggested outliers (or any others you select) in the same way as from the Local Model plots. Leave this one. If you zoom in on the plot (Shift-click-drag or middle-click-drag) you can see the value of the studentized residual of this point more clearly. Double-click to return to the previous view.

Note    Obviously you should never remove outliers as a matter of course. However, this tutorial is designed to show you how the toolbox helps you to do this when required. The default outlier selection criterion is a studentized residual greater than 3, to bring your attention to possible outliers, but you should never remove data without good reasons. Remove enough points and the model will simply interpolate the data and become useless for prediction. You can customize the criteria for outlier selection.

4. Select the other three response features in turn: max, Bhigh 2, and Blow 2. You will see that Blow 2 has a suggested outlier with a very large studentized residual; it is a good distance away from all the other data points for this response feature. All the other points are so clustered that removing this one could greatly improve the fit of the model to the remaining points, so remove it.

Return to the Local Model pane by clicking the local node PS22 in the Model Browser tree.

Verifying the Model

Selecting the Two-Stage Model