Skip to contents

Changelog

Source: NEWS.md

cjdiag 0.2.1

Focuses the package on a smaller set of well-supported plots and rewrites the README around the random-forest results table.

Breaking changes

  • Removed plot(tree_obj, type = "importance"). Decision trees are rendered via rpart.plot only; the bar chart of rpart variable importance was removed because the tree itself is the diagnostic and the bar chart was misleading at level resolution.
  • Removed plot(nmm_obj, type = "decisiveness") and plot(nmm_obj, type = "sample"). Nested marginal means now plots the cumulative-explanation curve only. The underlying decisiveness and sample_history columns are still available on the fit object.

Improvements

  • Cumulative and rank plots now place each label at its own rotated x-axis tick, so collisions are impossible by construction and no labels are dropped. Tick labels show the attribute on top in black and the level below in dark grey, separated by a line break, via ggtext::geom_richtext() (with a plain rotated-text fallback if ggtext is unavailable). The bottom plot margin scales dynamically with the longest label.
  • Decision tree plots replace the dummy-name labels (e.g. JobPlansno.plans.to.look.for.work) with attribute: level (e.g. JobPlans: no plans to look for work) via a custom split.fun driven by the fit’s attr_map. Works for any conjoint data without per-dataset configuration.
  • README: random-forest example shows the full results table via knitr::kable() with column-by-column explanations of mda, root_pct, class_0, and class_1. Importance plot switched from bar chart to the rank plot.
  • README and Getting Started vignette now include an Estimands Table (estimand / method = / question / output / behavioural assumption / when to use) so users can pick a method without reading the docs.
  • The single Introduction vignette is replaced by a Getting Started overview (cjdiag.Rmd) plus one task-oriented vignette per method (forest, tree, nmm, marginal_r2, crt).
  • Plot help pages now share an @family plotting block so the help pages cross-link.
  • README: added ERC AGAPP funding acknowledgement.

cjdiag 0.2.0

Initial CRAN release.

Methods

  • Random forest (method = "forest"): attribute importance via Mean Decrease in Accuracy, Mean Decrease in Gini, and root node appearance rates. Supports level-specific and attribute-level resolution.
  • Decision tree (method = "tree"): CART classification tree revealing hierarchical decision structure. Root split identifies the gatekeeper attribute.
  • CRT/HierNet (method = "crt"): L1-regularized logistic regression across a lambda grid. Lambda path, cross-validation, and permutation importance (MDA).
  • Nested marginal means (method = "nmm"): sequential elimination procedure. Decisiveness ranking reveals the decision order.
  • Marginal R-squared (method = "marginal_r2"): per-respondent per-attribute importance metric.

Features

  • Unified cj_fit(formula, data, method) interface for all methods.
  • importance() generic with methods for all model types.
  • Plot customization: palette, base_size, label_wrap, attribute.names, level.names, group_by_attribute, and full theme override.
  • Three palettes: "default", "colorblind" (Okabe-Ito), "grey".
  • Global options: set_cjdiag_theme(), set_cjdiag_labels(), get_cjdiag_options().
  • tidy() and glance() methods (conditional on broom/generics).
  • Bundled immig dataset (Hainmueller & Hopkins 2015).