Target Audience: Policy analysts, development researchers, program evaluators, and data practitioners working with complex social datasets
Prerequisites: Bivariate Analysis 101 or equivalent experience with correlations, cross-tabs, and basic statistical inference
Materials Needed: Statistical software (R/Stata/Python), notebooks, sample policy datasets, flip chart paper
By the end of this workshop, participants will be able to:
Simple Analysis: "Districts with better school infrastructure have 23% higher learning outcomes. Therefore, building better schools improves education."
The Problem: This ignores that districts with better infrastructure are also typically:
The Question: How much of the 23% difference is truly due to infrastructure vs. these other factors?
Observed Association: Women with microfinance loans score 40% higher on empowerment indices.
Possible Confounders:
Multivariate Approach: Control for baseline empowerment, location characteristics, other programs, and economic conditions to isolate the effect of microfinance specifically.
Every multivariate model has different types of variables serving different purposes in our causal story.
What you want to explain or predict
Examples: Learning outcomes, income, health status, program participation
Main variable of interest - your intervention or policy
Examples: Training program, policy change, infrastructure investment
Other factors that might influence Y - need to account for them
Examples: Age, education, income, location, baseline conditions
Variables that explain HOW X affects Y
Examples: Skills acquired (mediates training→income), knowledge (mediates education→behavior)
Variables that change the strength of X→Y relationship
Examples: Gender (training effects differ by gender), baseline poverty (program effects vary by initial wealth)
Variables that influence both X and Y, creating spurious association
Examples: Motivation affects both program participation and outcomes
Scenario: You're evaluating whether community health worker (CHW) programs improve child nutrition outcomes in rural Rajasthan.
Research Question: Do communities with CHW programs have better child nutrition outcomes than communities without such programs?
Activity Steps:
Key Questions to Consider:
Research Question: What is the return to education in Indian labor markets?
Model: Monthly Income = β₀ + β₁(Years of Education) + β₂(Experience) + β₃(Experience²) + β₄(Urban) + β₅(Male) + β₆(Caste) + ε
Sample Results:
Key Insight: The "8% return to education" controls for gender, location, experience, and caste - it's the pure effect of education, not confounded by these other differences.
Use when: Continuous outcome (income, test scores, BMI)
Interprets as: Unit change in Y per unit change in X
Example: "Each year of mother's education increases child height by 0.8 cm"
Use when: Binary outcome (employed/unemployed, program participation)
Interprets as: Odds ratios or predicted probabilities
Example: "Each year of education increases odds of formal employment by 1.15x"
Use when: Comparing group means while controlling for other variables
Interprets as: Adjusted group differences
Example: "Treatment group scores 12 points higher, adjusting for baseline scores"
Use when: Panel data with unobserved individual characteristics
Interprets as: Within-individual changes over time
Example: "Job training increases individual wages by 15% compared to their pre-training levels"
Critical Limitation: "Controlling for X" only works if you've measured X accurately and completely. Unmeasured confounders remain a major threat to causal interpretation in observational data.
Multivariate analysis is a powerful tool for testing causal hypotheses by controlling for alternative explanations. However, it's only as good as your theory, research design, and data quality. Use it to strengthen causal arguments, not to replace careful thinking about what causes what.
Essential Textbooks:
South Asian Development Context:
Practical Software Resources:
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 complete multivariate analysis datasets, code templates, and diagnostic tools, visit the ImpactMojo platform.
