Tutorial

This tutorial walks through a complete analysis using TSCSMethods.jl, explaining each step in detail.

Complete Workflow Overview

The following diagram shows the complete user workflow from data to results:

Each step is explained in detail below.

Understanding the Data Structure

TSCSMethods expects staggered treatment design data where:

• Units (e.g., counties, countries) are treated at specific times
• Treatment is binary (0/1) and occurs on specific dates, not continuously
• Time periods are relative to treatment: negative for pre-treatment, positive for post-treatment

Example Data Format

using TSCSMethods, DataFrames

# Generate example data
dat = example_data(n_units=20, n_days=50, seed=123)
first(dat, 10)

Key columns:

• fips: Unit identifier
• day: Time period (0-based)
• gub: Treatment indicator (1 only on treatment day)
• death_rte: Outcome variable
• pop_dens: Covariate

Step 1: Model Creation

The makemodel function sets up your causal inference model:

model = makemodel(
    dat,              # Your data
    :day,             # Time variable
    :fips,            # Unit identifier  
    :gub,             # Treatment variable
    :death_rte,       # Outcome variable
    [:pop_dens],      # Covariates for matching
    Dict(:pop_dens => false),  # Which covariates are time-varying
    5:10,             # F: Post-treatment periods to estimate
    -15:-10           # L: Pre-treatment periods for matching
)

Key Parameters Explained

• F periods (5:10): How many periods after treatment to estimate effects
• L periods (-15:-10): Which pre-treatment periods to use for matching
  • Must be negative for pre-treatment
  • Used to find similar control units
• Time-varying covariates: Set true if covariate changes over time

Step 2: Matching

Find control units similar to treated units in pre-treatment periods:

match!(model, dat)

# Check how many matches were found
println("Found matches for $(length(model.matches)) treated observations")

The matching algorithm:

1. Identifies treated units and their treatment times
2. Finds control units with similar covariate values in pre-treatment periods
3. Creates matched sets for each treated observation

Step 3: Balancing

Assess how well matching achieved covariate balance:

balance!(model, dat)

# Check balance results
model.meanbalances

Good balance means treated and control groups have similar covariate distributions in pre-treatment periods.

Step 4: Estimation

Estimate average treatment effects with bootstrap inference:

# Run estimation (without Bayesian factors for simplicity)
estimate!(model, dat; dobayesfactor=false)

# View results
model.results

Results include:

• f: Time periods relative to treatment
• att: Average treatment effect estimates
• q025, q975: 95% confidence intervals
• treated: Number of treated units
• matches: Number of control units matched

Interpreting Results

# Look at results
println("Treatment Effects by Time Period:")
select(model.results, :f, :att, :q025, :q975)

• Positive ATT: Treatment increased the outcome
• Negative ATT: Treatment decreased the outcome
• Confidence intervals: Statistical uncertainty around estimates
• Multiple time periods: See how effects evolve over time

Complete Example

using TSCSMethods, DataFrames

# 1. Load/generate data
dat = example_data(n_units=30, n_days=60, seed=42)

# 2. Create model  
model = makemodel(dat, :day, :fips, :gub, :death_rte, 
                 [:pop_dens], Dict(:pop_dens => false),
                 3:8, -20:-10)

# 3. Run complete workflow
match!(model, dat)
balance!(model, dat) 
estimate!(model, dat; dobayesfactor=false)

# 4. Examine results
println("Number of treated observations: ", model.treatednum)
println("Average treatment effect in period 3: ", model.results[1, :att])