Target Audience: Development economists, policy analysts, impact evaluators, researchers, and data scientists working on causal questions
Prerequisites: Multivariate Analysis 101 or equivalent regression knowledge, basic statistics
Materials Needed: Statistical software (R/Stata/Python), sample datasets, calculators
By the end of this workshop, participants will be able to:
Observed Correlation: People with more education earn more money (r = 0.6 in Indian NSS data)
But is it causal? Consider these alternative explanations:
The Policy Question: If we expand access to education, will incomes actually rise?
"No causation without manipulation" - To establish causation, we need to compare what happens to the same unit when it receives treatment versus when it doesn't. But we can never observe both states for the same person at the same time.
The Counterfactual: What would have happened to the treated units if they had not been treated?
For each individual i:
Individual Treatment Effect: δᵢ = Y₁ᵢ - Y₀ᵢ
The Problem: We only observe one potential outcome per person!
Average Treatment Effect (ATE): δ = E[Y₁ᵢ - Y₀ᵢ] = E[Y₁ᵢ] - E[Y₀ᵢ]
Naive Comparison: E[Y|T=1] - E[Y|T=0] = ATE + Selection Bias
Problem: Third variables affect both treatment and outcome
Example: Family wealth affects both education access and income opportunities
Solution: Control for confounders in regression
Problem: Outcome affects treatment, not just vice versa
Example: Health affects income, but income also affects health
Solution: Instrumental variables, panel data
Problem: Treatment groups differ systematically
Example: Motivated students choose higher education
Solution: Random assignment, natural experiments
Problem: Imperfect measurement of variables
Example: Self-reported income vs actual income
Solution: Better data, instrumental variables
Scenario: A study finds that microfinance borrowers have higher business profits than non-borrowers.
Your Task: Identify potential sources of bias in concluding that microfinance causes higher profits:
What variables might affect both microfinance access and profits?
How might microfinance borrowers differ systematically?
Could higher profits lead to microfinance access?
What measurement problems might exist?
Key Assumption: Conditional on observed covariates X, treatment assignment is "as good as random"
Causal Interpretation: β₁ is the average treatment effect IF we control for all confounders
The Big IF: This requires controlling for ALL variables that affect both treatment and outcome
1. Theory First: Draw a causal diagram (DAG) showing relationships
2. Include Confounders: Control for variables that affect both T and Y
3. Avoid Bad Controls: Don't control for post-treatment variables
4. Check Robustness: See if results change with different specifications
Research Question: Does participating in an employment training program increase wages?
Available Data:
| Variable | Include as Control? | Justification |
|---|---|---|
| Age | Yes | Affects both training participation and wages |
| Education | Yes | Confounder - affects participation and outcomes |
| Previous experience | Yes | Pre-treatment variable affecting both |
| Motivation score | Yes | Affects who joins training and wage outcomes |
| Job search behavior | No | Post-treatment - might be mediator |
| Regional unemployment | Yes | Affects labor market opportunities |
Example Regression Output:
Interpretation:
When regression isn't enough, consider these approaches:
| Method | When to Use | Key Assumption | Example |
|---|---|---|---|
| Instrumental Variables | Endogenous treatment, unobserved confounders | Valid instruments exist | Distance to school as IV for education |
| Difference-in-Differences | Policy changes, panel data | Parallel trends | Minimum wage policy impact |
| Regression Discontinuity | Arbitrary assignment thresholds | No manipulation of running variable | Scholarship eligibility based on test scores |
| Fixed Effects | Unobserved time-invariant factors | No time-varying confounders | Individual or location fixed effects |
| Matching | Observational data, selection on observables | Conditional independence | Propensity score matching |
Method choice depends on:
There's no universally best method - the right choice depends on context and credible identification strategy.
Foundational Texts:
Applied Development Economics:
Software and Tools:
Data Sources for Practice:
Next Steps in ImpactMojo:
This handout is part of the ImpactMojo 101 Knowledge Series
Licensed under CC BY-NC-ND 4.0 • Free to use with attribution • www.impactmojo.in
For regression templates, causal inference checklists, and analysis code examples, visit the ImpactMojo platform.
