Accuracy and classification results (generalized linear mixed models)

Figure 1. Accuracy chart from Model Summary view of generalized logit mixed model
Accuracy chart from Model Summary view of generalized logit mixed model
  1. Activate (double-click) the model object for the generalized logit mixed model.

    Since the target is categorical, the Model Summary view includes a chart of the overall model accuracy.

    Figure 2. Accuracy chart from Model Summary view of generalized logit model
    Accuracy chart from Model Summary view of generalized logit model
  2. Activate (double-click) the model object for the generalized logit mixed model.

    Comparing the accuracy charts tells you at a glance that the overall correct classification rate for the mixed model is far superior; 85.2% compared to 46.2%. For further details on how the model accuracies differ, see the classification tables.

    Figure 3. Classification view of generalized logit mixed model
    Classification view of generalized logit mixed model
  3. In the viewer for the generalized logit mixed model, click the Classification view thumbnail.

    For each case, the predicted response category is chosen by selecting the category with the highest model-predicted probability.

    • Cells on the diagonal are correct predictions.
    • Cells off the diagonal are incorrect predictions.

    This model does very well at identifying customers who have service, either with the company or with another provider, and adequately at identifying customers with no service. Based on the imbalance between the upper right and lower left off-diagonal cells in the table, when the model errs, it tends to overestimate the probability that a customer has service, especially with the company.

    Figure 4. Classification view of generalized logit model
    Classification view of generalized logit model
  4. In the viewer for the generalized logit model, click the Classification view thumbnail.

The model without random effects fails to identify any of the No service customers and correctly identifies only a third of the Other provider customers. This suggests that the generalized logit mixed model should be used.

Next