CHAID and CRT decision trees in your browser

Crosstabs runs the SPSS TREE procedure with three growing methods: CHAID and Exhaustive CHAID, which make multiway splits backed by Bonferroni-adjusted tests, CRT, which makes binary splits by impurity reduction and can be pruned, and QUEST, which picks the split variable with ANOVA F, chi-square and Levene tests and places binary splits by quadratic discriminant analysis.

• Which predictor separates the target best at each step, and which categories or value ranges belong together (the split values on every branch).
• The profile of every node: the target distribution or mean, its size and the predicted category.
• Which terminal nodes concentrate the target category, through Gain %, Response % and Index % in the gains table.
• How often the tree misclassifies cases, by resubstitution and, optionally, V-fold cross-validation.
• For CRT, how important each predictor is, counting both primary and surrogate splits.

1. 01

   Open the decision tree panel

   Open your file in the workspace, switch to Statistics mode and choose Decision tree (TREE) in the Classify group. The data stays in your browser.

2. 02

   Set the dependent variable and its level

   Choose the Dependent variable and confirm its Measurement level (Nominal, Ordinal or Scale); the level decides which tests and impurity measures are offered.

3. 03

   Tick the independent variables

   Tick the Independent variables and set each one's measurement level in the dropdown beside it. Scale predictors are binned for CHAID and split at thresholds for CRT.

4. 04

   Pick the growing method and stopping rules

   Under Growing method and criteria choose CHAID, Exhaustive CHAID or CRT, then set Maximum tree depth (3 for CHAID, 5 for CRT by default) and the minimum cases in parent and child nodes.

5. 05

   Tune the method-specific criteria

   For CHAID set the significance levels for splitting and merging and, for a nominal target, the Pearson or likelihood-ratio chi-square. For CRT pick the impurity measure and minimum change in improvement, and optionally tick Prune tree to avoid overfitting with a standard-error rule.

6. 06

   Validate and run

   Tick Cross-validation risk estimate and choose the sample folds, pick the target category for gains, then run. A workspace weight can be used only after you declare it a frequency weight.

Model Summary

Growing method, the independent variables that actually entered the tree, number of nodes and terminal nodes, depth and cases used, with a note on excluded cases.

Tree diagram

An interactive tree: each node shows its distribution or mean and N, each branch the predictor and split values; click a node to collapse its branch.

Tree Table

One row per node with the category percentages (or mean and standard deviation), N, predicted category, parent node, splitting variable, the CHAID statistic with its Bonferroni-adjusted p value or the CRT improvement, and the split values.

Gains for Nodes

Terminal nodes sorted by response: Gain % is the share of all target cases a node captures, Response % the rate inside it and Index % that rate relative to the whole sample; cumulative columns build the gains chart. A scale target gives the Gain Summary for Nodes by mean.

Risk

The misclassification rate (or within-node variance for a scale target) with its standard error, by resubstitution and by cross-validation. A large gap between the two signals overfitting.

Classification

Observed by predicted categories with percent correct per row and overall; CRT trees add Independent Variable Importance and, when pruned, the Cost-Complexity Pruning Sequence.