Visualizing Causal Assignment Trees for CSDiD and DR-DDD Designs

Tools for constructing, labeling, and visualizing Causal Assignment Trees (CATs) in settings with staggered adoption. Supports Callaway and Sant'Anna difference-in-differences (CSDiD) and doubly robust difference-in-difference-differences (DR-DDD) designs. The package helps clarify treatment timing, never-treated vs. not-yet-treated composition, and subgroup structure, and produces publication-quality diagrams and summary tables. Current functionality focuses on data-to-node mapping, node counts, cohort-year summaries, and high-quality tree plots suitable for empirical applications prior to estimation. Methods are based on Callaway and Sant'Anna (2021) , Sant'Anna and Zhao (2020) , and Kilanko (2026) < https://github.com/VictorKilanko/catviz>.

catviz: Causal Assignment Tree Visualization for Staggered DiD, DDD, and Related Designs

catviz is an R package for visualizing and understanding Causal Assignment Trees (CATs) — hierarchical structures that summarize treatment timing, subgroup composition, and sample classification in staggered difference-in-differences (CSDID) and staggered DDD like DRDDD and related causal inference frameworks.

It provides a publication-ready visualization of treated, control, and never-treated groups, along with counts and subgroup summaries, to help researchers verify sample balance and treatment assignment logic.

Working Example

The example creates simulated panel data for hospitals nested within states,
where states adopt treatment at different years, and hospitals may also belong
to binary subgroups (for DR-DDD analysis).

Variable definitions

Example code

# =======================================================
# Example: State-level staggered adoption with subgroups
# =======================================================

# Install if needed
# install.packages("devtools")
# devtools::install_github("VictorKilanko/catviz")

library(catviz)
library(dplyr)
library(tidyr)
library(purrr)  # for map()

set.seed(123)

# =======================================================
# 1. Define simulation setup
# =======================================================
states <- sprintf("S%02d", 1:20)    # 20 states
years  <- 2014:2023
N_hosp <- 5                         # 5 hospitals per state

# Assign first treatment year (g) per state
adopt_years <- c(2015, 2016, 2019, 2020, 2021, 2023, Inf)

state_level <- tibble(
  state = states,
  g = sample(adopt_years, length(states), replace = TRUE)
)

# =======================================================
# 2. Create hospitals nested within states
# =======================================================
hospitals <- state_level %>%
  mutate(
    hospital_id = map(state, ~ paste0(.x, "_H", 1:N_hosp))
  ) %>%
  unnest(hospital_id) %>%
  mutate(
    p = sample(0:1, n(), replace = TRUE)  # subgroup (omit for CSDID)
  )

# =======================================================
# 3. Expand to panel structure
# =======================================================
example_data <- expand_grid(
  hospital_id = hospitals$hospital_id,
  year = years
) %>%
  left_join(hospitals, by = "hospital_id") %>%
  arrange(hospital_id, year)

# =======================================================
# 4. Define CAT specification
# =======================================================
# Variables in the CSDID / DR-DDD framework:
# - id: unit of analysis (hospital_id)
# - group_id: grouping or treatment level (state)
# - time: time variable (year)
# - g: first treatment year for the group (state)
# - subgroup: subgroup classification (p), used for DR-DDD only

spec <- cat_spec(
  data      = example_data,
  id        = "hospital_id",
  time      = "year",
  g         = "g",
  subgroup  = "p",         # omit for pure CSDID
  group_id  = "state"      # treatment assigned at state level
)

# Label nodes for clarity
spec <- cat_label(spec)

# =======================================================
# 5. Summaries
# =======================================================
cat_counts(spec)   # counts per node (unit-level by default)

# =======================================================
# 6. Visualization
# =======================================================
dir.create("man/figures", recursive = TRUE, showWarnings = FALSE)

# Example 1: Default (unit-level counts)
out_units <- cat_plot_tree(
  spec,
  counts    = TRUE,
  count_by  = "units",   # counts unique hospitals
  save_plot  = "man/figures/CAT_plot_units.png",
  save_table = "man/figures/CAT_summary_units.csv"
)

# Example 2: Observation-level counts
out_obs <- cat_plot_tree(
  spec,
  counts    = TRUE,
  count_by  = "obs",     # counts hospital-year observations
  save_plot  = "man/figures/CAT_plot_obs.png",
  save_table = "man/figures/CAT_summary_obs.csv"
)

# Example 3: Hide counts (just structure)
out_nolabel <- cat_plot_tree(
  spec,
  counts    = FALSE,
  save_plot  = "man/figures/CAT_plot_nolabel.png"
)

# =======================================================
# 7. Display example plot
# =======================================================
print(out_units$plot)

# =======================================================
# 8. Confirm saved outputs
# =======================================================
message("Unit-level plot: man/figures/CAT_plot_units.png")
message("Observation-level plot: man/figures/CAT_plot_obs.png")
message("Summary tables saved in man/figures/")

Example outputs

1. CAT plot (unit-level counts)

Below is the automatically generated Causal Assignment Tree (CAT) showing treated, control, and never-treated branches, with counts based on unique units (hospitals).

• All Groups (n=100): Total number of unique hospitals
• Treated Groups (n=90): Hospitals in states that adopted treatment
• Never-Treated (g=∞, n=10): Hospitals in states that never adopted treatment
• Branches such as (3) t≥g, p=0 show subgroup and timing splits within treated groups

2. CAT plot (observation-level counts)

In this version, node counts reflect total observations (hospital-year combinations), not unique units.

This helps assess data coverage across pre- and post-treatment periods.

3. CAT structure only (no counts)

For schematic or publication purposes, you can hide counts entirely:

4. Treatment-year summary

The accompanying table summarizes the number of treated units by first treatment year and subgroup.

📁 The table is also saved automatically as:

• man/figures/CAT_summary_units.csv
• man/figures/CAT_summary_obs.csv

Interpreting the CAT visualization

The Causal Assignment Tree decomposes the dataset into mutually exclusive groups based on:

1. Treatment timing (g)
2. Pre/post period (t < g vs t ≥ g)
3. Subgroup (p)

Each node in the tree represents a distinct subset of the data, and the associated count (n) corresponds to the number of unique hospitals (or observations, depending on the option selected).

Reading the branches

• All Groups: The entire dataset of hospitals.
• Treated Groups: States that eventually receive treatment.
• Never-Treated (g=∞): States that never receive treatment.
• t < g: Pre-treatment observations (before adoption).
• t ≥ g: Post-treatment observations (after adoption).
• p = 0 / 1: Subgroup categories (for DR-DDD).

Example interpretation

• 90 hospitals belong to treated states; 10 are never treated.
• Among treated hospitals:
  • Those observed before their treatment year (t < g) appear under (1) and (2).
  • Those observed after treatment (t ≥ g) appear under (3) and (4).
• The subgroup split (p=0, p=1) reveals balance across treated vs control subpopulations.

5. Why count type matters

By default, cat_plot_tree() counts unique units (count_by = "units"), which is consistent with CSDID or DR-DDD analysis where the treatment effect is at the unit level.
However, users can also choose count_by = "obs" to count total unit-year observations, which helps verify panel balance or data coverage.

Output verification

All outputs are saved in the man/figures/ directory: