#805: Add node p-value display to plot_causal_tree

[genrichez]

Proposed changes

Add the ability to display treatment effect p-values at each node when plotting a causal tree via plot_causal_tree. This addresses #805.

When node_pvalues=True is passed to CausalTreeRegressor during fit, a Welch's t-test is computed at each node comparing treatment vs control outcomes. The resulting p-values can then be rendered in the plot by passing pvalue=True to plot_causal_tree.

Usage:

model = CausalTreeRegressor(control_name=0, groups_cnt=True, node_pvalues=True)
model.fit(X=X, treatment=treatment, y=y)
plot_causal_tree(model, pvalue=True)

Each node will display p_value(group) = X.XXXX for each treatment group. Nodes with fewer than 2 samples in either group show N/A.

Types of changes

• Bugfix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation Update (if none of the other choices apply)

Checklist

• I have read the CONTRIBUTING doc
• I have signed the CLA
• Lint and unit tests pass locally with my changes
• I have added tests that prove my fix is effective or that my feature works
• I have added necessary documentation (if appropriate)
• Any dependent changes have been merged and published in downstream modules

Further comments

Implementation follows the same pattern as the existing groups_cnt feature:

• node_pvalues=True in the constructor enables computation during fit()
• _compute_node_pvalues() traverses the tree via decision_path (same approach as _count_groups_distribution)
• Uses Welch's t-test (scipy.stats.ttest_ind, equal_var=False) which is standard for unequal group sizes
• Multi-treatment support: computes a separate p-value for each treatment group vs control
• Gracefully handles nodes with insufficient samples (returns None)

This PR was co-authored with AI assistance.

[CLAassistant]

All committers have signed the CLA.

[jeongyoonlee]

The node p-values are naive in-sample within-node t-tests — the partition is learned from the same data and splits maximize effect separation, so they're anti-conservative (no honesty/selective-inference or multiple-testing correction) and are raw diff-in-means (valid only under in-node unconfoundedness). Fine as the diagnostic #805 asks for, but please add a docstring note on node_pvalues / plot_causal_tree(pvalue=...) that these are descriptive, not valid post-selection inference.

Minor: a zero-variance node makes ttest_ind return nan, rendered as p_value(g) = nan — worth mapping to N/A like the insufficient-samples branch.

[genrichez]

The node p-values are naive in-sample within-node t-tests — the partition is learned from the same data and splits maximize effect separation, so they're anti-conservative (no honesty/selective-inference or multiple-testing correction) and are raw diff-in-means (valid only under in-node unconfoundedness). Fine as the diagnostic #805 asks for, but please add a docstring note on node_pvalues / plot_causal_tree(pvalue=...) that these are descriptive, not valid post-selection inference.

Minor: a zero-variance node makes ttest_ind return nan, rendered as p_value(g) = nan — worth mapping to N/A like the insufficient-samples branch.

Thanks for the thorough review! I've addressed these.
I've added the cautionary docstring notes for node_pvalues in both CausalTreeRegressor and plot_causal_tree, clarifying that these are descriptive p-values and not for post-selection inference ,also implemented the NaN to N/A mapping for p-values in the plot, and added a regression test to ensure this behavior is consistent.