Visual Six Sigma: Making Data Analysis Lean

9.5. Modeling Relationships: Logistic Model

Recall that Mi-Ling intends to explore three modeling approaches: logistic regression, recursive partitioning, and neural nets. She considers using discriminant analysis, but recalls that discriminant analysis requires assumptions about the predictor variables—for each group, the predictors must have a multivariate normal distribution, and unless one uses quadratic discriminant analysis, the groups must have a common covariance structure. Thinking ahead to her test campaign data, Mi-Ling notes that some of her predictors will be nominal. Consequently, discriminant analysis will not be appropriate, whereas logistic regression will be a legitimate approach.

Mi-Ling begins her modeling efforts with a traditional logistic model. But before launching into a serious logistic modeling effort, Mi-Ling decides that she wants to see how a logistic model, based on only two predictors, might look. She begins by fitting a model to only two predictors that she chooses rather haphazardly.

Once she has visualized a logistic fit, she proceeds to fit a logistic model that considers all 30 of her variables as candidate predictors. She uses a stepwise regression procedure to reduce this model.

9.5.1. Visualization of a Two-Predictor Model

Since she will use only the training data in constructing her models, Mi-Ling copies the values in the row state variable Training Set to the row states in the data table CellClassification_2.jmp. As she did earlier, ...