feat(chaid): emit split test stats (chi-square/p-value/df) in tree_nodes

R/build_chaid.R

@@ -429,6 +429,10 @@ build_chaid_tree_nodes <- function(x) {
     })
     class_json <- as.character(jsonlite::toJSON(arr, auto_unbox = TRUE, digits = NA))
 
+    # Split-test stats belong to nodes that actually split; NA everywhere else
+    # (leaves, and nodes whose best candidate failed the alpha_split gate).
+    is_split <- !isTRUE(nodes$is_terminal[i])
+
     edge_row <- which(edges$child_id == id)
     if (length(edge_row) == 1) {
       cond_column <- map_name(edges$split_variable[edge_row])
@@ -458,6 +462,12 @@ build_chaid_tree_nodes <- function(x) {
       sd_value = NA_real_,
       rmse_value = NA_real_,
       dist_json = NA_character_,
+      # CHAID split test at this node (NA for leaves). The interactive tree
+      # renderer shows these on the splitting (parent) node, SPSS-style.
+      p_value = if (is_split) nodes$p_value[i] else NA_real_,
+      adjusted_p_value = if (is_split) nodes$adjusted_p_value[i] else NA_real_,
+      split_statistic = if (is_split) nodes$split_statistic[i] else NA_real_,
+      split_df = if (is_split) nodes$split_df[i] else NA_real_,
       stringsAsFactors = FALSE
     )
   })

R/chaid.R

@@ -86,6 +86,8 @@ chaid_fit <- function(data,
     split_variable = NA_character_,
     p_value = NA_real_,
     adjusted_p_value = NA_real_,
+    split_statistic = NA_real_,
+    split_df = NA_real_,
     rule = 'Root',
     stringsAsFactors = FALSE
   )
@@ -655,6 +657,8 @@ grow_chaid_tree <- function(data, target, predictors, parameters,
       split_variable = NA_character_,
       p_value = NA_real_,
       adjusted_p_value = NA_real_,
+      split_statistic = NA_real_,
+      split_df = NA_real_,
       rule = rule
     )
     state$node_records[[as.character(node.id)]] <- record
@@ -705,6 +709,8 @@ grow_chaid_tree <- function(data, target, predictors, parameters,
     current.record <- state$node_records[[as.character(node.id)]]
     current.record$p_value <- best$p_value
     current.record$adjusted_p_value <- best$adjusted_p_value
+    current.record$split_statistic <- best$test$statistic
+    current.record$split_df <- best$test$df
     state$node_records[[as.character(node.id)]] <- current.record
 
     if (is.na(best$p_value) || best$adjusted_p_value > parameters$alpha_split) {
@@ -769,6 +775,8 @@ grow_chaid_tree <- function(data, target, predictors, parameters,
       split_variable = record$split_variable,
       p_value = record$p_value,
       adjusted_p_value = record$adjusted_p_value,
+      split_statistic = record$split_statistic,
+      split_df = record$split_df,
       rule = record$rule,
       stringsAsFactors = FALSE
     )

tests/testthat/test_build_chaid.R

@@ -66,6 +66,35 @@ test_that("tree_nodes tidy output matches the renderer schema", {
   expect_true(all(c("label", "n", "pct") %in% names(parsed)))
 })
 
+test_that("tree_nodes carries CHAID split stats on split nodes, NA on leaves", {
+  # Build a target that genuinely depends on `plan` so CHAID makes a split.
+  set.seed(3)
+  n <- 400
+  plan <- sample(c("A", "B"), n, replace = TRUE)
+  is_churn <- ifelse(plan == "A", runif(n) < 0.8, runif(n) < 0.2)
+  df <- data.frame(is_churn = is_churn, plan = plan,
+                   region = sample(c("east", "west"), n, replace = TRUE),
+                   stringsAsFactors = FALSE)
+  model_df <- suppressWarnings(exp_chaid(df, is_churn, plan, region,
+                                         min_split = 20, min_bucket = 5))
+  nodes <- model_df %>% tidy_rowwise(model, type = "tree_nodes")
+
+  stat_cols <- c("p_value", "adjusted_p_value", "split_statistic", "split_df")
+  expect_true(all(stat_cols %in% colnames(nodes)))
+
+  split_nodes <- nodes[!nodes$is_leaf, ]
+  leaf_nodes <- nodes[nodes$is_leaf, ]
+  expect_gt(nrow(split_nodes), 0)
+  # Every splitting node reports a finite chi-square, df, and adjusted p-value.
+  expect_true(all(is.finite(split_nodes$split_statistic)))
+  expect_true(all(is.finite(split_nodes$split_df)))
+  expect_true(all(is.finite(split_nodes$adjusted_p_value)))
+  expect_true(all(split_nodes$adjusted_p_value >= 0 & split_nodes$adjusted_p_value <= 1))
+  # Leaves never carry a split test.
+  expect_true(all(is.na(leaf_nodes$split_statistic)))
+  expect_true(all(is.na(leaf_nodes$split_df)))
+})
+
 test_that("binary tree_nodes lists the positive (TRUE) class first", {
   df <- make_binary_df()
   model_df <- suppressWarnings(exp_chaid(df, is_churn, plan, region,