sdid 0.2.0

Synthetic Difference-in-Differences (SDID) for causal inference in Python

Synthetic Difference-in-Differences (SDID)

English | 繁體中文

---

English

A Python implementation of Synthetic Difference-in-Differences for causal inference.

Overview

SDID combines synthetic control methods with difference-in-differences to provide robust causal effect estimates. It automatically finds optimal weights for:

• Control units: Creates a synthetic comparison group
• Time periods: Balances pre/post treatment comparisons

Installation

From PyPI (Recommended)

pip install sdid

From Source

git clone https://github.com/yourusername/sdid.git
cd sdid
pip install .

Development Installation

git clone https://github.com/yourusername/sdid.git
cd sdid
uv sync --extra dev

Quick Start

import pandas as pd
from sdid import SyntheticDiffInDiff

# Load your panel data
data = pd.read_csv("your_data.csv")

# Initialize and fit
sdid = SyntheticDiffInDiff(
    data=data,
    outcome_col="outcome",
    times_col="year",
    units_col="state",
    treat_col="treated",
    post_col="post"
)

# Estimate treatment effect
effect = sdid.fit()
print(f"Treatment Effect: {effect:.4f}")

# Estimate standard error
sdid.estimate_se(n_bootstrap=200, n_jobs=4)
print(sdid.summary())

Data Format

Your data should be in long format (one row per unit-time observation):

Column	Description	Example
outcome	Outcome variable	Sales, GDP
times	Time period	2018, 2019, 2020
units	Unit identifier	"CA", "TX", "NY"
treat	Treatment indicator	0 = control, 1 = treated
post	Post-treatment period	0 = before, 1 = after

Example:

unit  year  outcome  treated  post
CA    2018  100.5    0        0
CA    2019  102.3    0        0
CA    2020  105.1    0        1
TX    2018  98.2     1        0
TX    2019  99.8     1        0
TX    2020  108.7    1        1

API Reference

Method	Description
fit(verbose=False)	Fit model and return treatment effect
estimate_se(n_bootstrap, seed, n_jobs)	Estimate standard error via placebo bootstrap
summary()	Return formatted results summary
run_event_study(times)	Estimate effects for multiple periods
plot_event_study(times, ...)	Create event study plot with confidence intervals
get_weights_summary()	Return unit and time weights
is_fitted	Property to check if model is fitted

Event Study

# Estimate dynamic treatment effects
effects = sdid.run_event_study([2020, 2021, 2022])

# Plot with confidence intervals
fig = sdid.plot_event_study(
    times=[2020, 2021, 2022],
    n_bootstrap=200,
    confidence_level=0.95,
    n_jobs=4
)
fig.savefig("event_study.png", dpi=300)

Assumptions

SDID relies on these key assumptions:

1. No anticipation: Units don't change behavior before treatment
2. SUTVA: No spillover effects between units
3. Overlap: Control units can approximate treated units

Reference

@article{arkhangelsky2021synthetic,
  title={Synthetic difference-in-differences},
  author={Arkhangelsky, Dmitry and Athey, Susan and Hirshberg, David A and Imbens, Guido W and Wager, Stefan},
  journal={American Economic Review},
  volume={111},
  number={12},
  pages={4088--4118},
  year={2021}
}

Development

# Install dev dependencies
uv sync --extra dev

# Run tests
uv run pytest tests/ -v

# Run linter
uv run ruff check .

# Format code
uv run ruff format .

# Build package
uv build

# Upload to PyPI (requires API token)
uv publish

License

MIT License - see LICENSE for details.

---

繁體中文

合成雙重差分法 (SDID) 的 Python 實現，用於因果推論。

概述

SDID 結合了合成控制法與雙重差分法的優點，提供穩健的因果效應估計。它自動計算最佳權重：

• 控制單位權重：建立合成對照組
• 時間期間權重：平衡處理前後的比較

安裝

從 PyPI 安裝（推薦）

pip install sdid

從原始碼安裝

git clone https://github.com/yourusername/sdid.git
cd sdid
pip install .

開發安裝

git clone https://github.com/yourusername/sdid.git
cd sdid
uv sync --extra dev

快速開始

import pandas as pd
from sdid import SyntheticDiffInDiff

# 載入面板資料
data = pd.read_csv("your_data.csv")

# 初始化並擬合
sdid = SyntheticDiffInDiff(
    data=data,
    outcome_col="outcome",    # 結果變數
    times_col="year",         # 時間
    units_col="state",        # 單位識別
    treat_col="treated",      # 處理指標
    post_col="post"           # 處理後指標
)

# 估計處理效果
effect = sdid.fit()
print(f"處理效果: {effect:.4f}")

# 估計標準誤
sdid.estimate_se(n_bootstrap=200, n_jobs=4)
print(sdid.summary())

資料格式

資料須為長格式（每個單位-時間組合一列）：

欄位	說明	範例
outcome	結果變數	銷售額、GDP
times	時間期間	2018, 2019, 2020
units	單位識別	"CA", "TX", "NY"
treat	處理指標	0 = 控制組, 1 = 處理組
post	處理後期間	0 = 處理前, 1 = 處理後

範例：

unit  year  outcome  treated  post
CA    2018  100.5    0        0
CA    2019  102.3    0        0
CA    2020  105.1    0        1
TX    2018  98.2     1        0
TX    2019  99.8     1        0
TX    2020  108.7    1        1

API 參考

方法	說明
fit(verbose=False)	擬合模型並回傳處理效果
estimate_se(n_bootstrap, seed, n_jobs)	透過安慰劑 bootstrap 估計標準誤
summary()	回傳格式化的結果摘要
run_event_study(times)	估計多個時間點的效果
plot_event_study(times, ...)	繪製帶信賴區間的事件研究圖
get_weights_summary()	回傳單位和時間權重
is_fitted	檢查模型是否已擬合的屬性

事件研究

# 估計動態處理效果
effects = sdid.run_event_study([2020, 2021, 2022])

# 繪製帶信賴區間的圖表
fig = sdid.plot_event_study(
    times=[2020, 2021, 2022],
    n_bootstrap=200,
    confidence_level=0.95,
    n_jobs=4
)
fig.savefig("event_study.png", dpi=300)

假設條件

SDID 依賴以下關鍵假設：

1. 無預期效應：單位不會在處理前改變行為
2. SUTVA：單位間無外溢效應
3. 重疊性：控制單位能近似處理單位

參考文獻

@article{arkhangelsky2021synthetic,
  title={Synthetic difference-in-differences},
  author={Arkhangelsky, Dmitry and Athey, Susan and Hirshberg, David A and Imbens, Guido W and Wager, Stefan},
  journal={American Economic Review},
  volume={111},
  number={12},
  pages={4088--4118},
  year={2021}
}

開發

# 安裝開發相依套件
uv sync --extra dev

# 執行測試
uv run pytest tests/ -v

# 執行 linter
uv run ruff check .

# 格式化程式碼
uv run ruff format .

# 建置套件
uv build

# 上傳到 PyPI（需要 API token）
uv publish

授權

MIT 授權 - 詳見 LICENSE。