MATCHING METHODS

MATCHING METHODS

Introduction

  • Definition: Matching methods compare individuals who received a treatment (like a program) with those who didn’t to assess treatment effects.

  • Non-parametric Nature: No strict assumptions about treatment effects make matching methods flexible compared to others.

Advantages

  • Less Specification: Requires fewer assumptions than regression-based approaches, useful when random experimentation isn't possible.

  • Estimate Causal Effects: Allows researchers to estimate how one thing affects another without needing randomized experiments by using observational data.

Key Assumptions of Matching Estimators

  1. Independence of Treatment: The outcome depends only on the treatment, not on unobservable factors.

  2. No Unobserved Factors/Confounders: No hidden confounders should influence both the treatment and outcomes to ensure accurate outcomes.

  3. Comparable Groups: Treatment and non-treatment groups should be similar based on observable characteristics, ensuring fairness and unbiased results (probability between 1 and 0).

  4. Selection on Unobservables: Should not affect results.

Balancing Requirement

  • Balancing Method: Ensures treated and untreated groups have similar pre-treatment characteristics for valid comparisons.

  • Propensity Score: Utilizes the probability of receiving treatment to check if groups are matched in terms of characteristics.

Limitations

  • Hard to Balance: Difficult to ensure balanced pre-treatment characteristics can lead to bias in results.

Matching Estimator with Common Support

  • Understanding Common Support: This relates to 'overlap,' ensuring groups are not too different for effective comparison.

  • Validation: Comparison is only valid if groups overlap.

Propensity Score Matching (PSM)

  • Definition: Balancing requirement between treated and untreated groups based on scores from observable characteristics (like income, age).

  • Comparison: Individuals with similar propensity scores are compared.

Matching Estimators of the ATT Based on the Propensity Score

  • Simply estimating the propensity score is insufficient for Average Treatment Effect on the Treated (ATT) estimation.

  • Four Widely Used Matching Methods:

    1. Nearest Neighbor Matching: Closest match based on scores.

    2. Radius Matching: Limited range of scores.

    3. Kernel Matching: Uses weighted average based on scores.

    4. Stratification Matching: Groups individuals into score intervals.

Role of Participation Decision

  • Observable Information: Uses observable characteristics (like income, age) to predict likelihood of joining treatment programs.

Drawbacks of Matching Estimators

  1. Finding the Right Counterfactual: Difficulty in finding similar individuals for treatment comparisons.

  2. Limited Common Support: Lack of overlap between treated and non-treated individuals complicates comparisons.

  3. Heavy Data Requirements: Large datasets are essential for effective matching, which can be a significant limitation.

  4. Variable Selection for Propensity Score: Choosing which variables to include can affect results.

Solutions

  • Sensitivity Analysis: Tests robustness of results by varying assumptions or including different variables.

  • Best Practices: Match participants within the same geographical area and using identical survey tools to enhance comparability of outcomes.

Combining Matching with Other Methods

  • Difference-in-Difference (DiD): Combines matching with DiD to estimate treatment effects using panel data, comparing changes over time between treated and non-treated groups.

  • Addressing Selection Bias: Matching estimators create a more comparable control group while DiD manages time-varying confounders.

Advantages of Combining Methods

  • Improved Comparisons: Enhances robustness of results by addressing biases from multiple angles.