Understanding Causal Inference: Key Concepts Explained

Causal Inference is the process of determining whether a cause-and-effect relationship exists between two variables, beyond mere correlation. It aims to answer questions like “Does X actually cause Y?” rather than just observing that “X and Y are associated.”

Key Concepts in Causal Inference:

1. Association vs. Causation
   • Association (Correlation): Two variables change together (e.g., ice cream sales and drowning incidents both rise in summer).
   • Causation: One variable directly influences the other (e.g., smoking → lung cancer).
2. Confounding Variables
   • A third variable that affects both the supposed cause and effect (e.g., heat in the ice cream/drowning example).
   • Proper causal inference requires controlling for confounders.
3. Counterfactuals
   • The idea of comparing what happened with what would have happened if the cause had not occurred (e.g., “Would this patient have recovered without the drug?”).
4. Experimental vs. Observational Data
   • Experimental: Randomized Controlled Trials (RCTs) (e.g., A/B tests) allow strong causal claims.
   • Observational: Real-world data (e.g., surveys, medical records) require advanced methods to infer causality.

Methods for Causal Inference:

• Randomized Experiments (RCTs): Gold standard (random assignment eliminates confounders).
• Matching: Pairing similar subjects (e.g., smokers/non-smokers with matching age, health).
• Regression Adjustment: Statistically controlling for confounders.
• Instrumental Variables (IV): Using an external factor that affects only the cause.
• Difference-in-Differences: Comparing changes over time between treated/untreated groups.
• Causal Graphs (DAGs): Diagrams to model relationships and identify confounders.

Example Applications:

• Medicine: Does a new drug improve patient outcomes?
• Economics: Does education increase earnings?
• Marketing: Did an ad campaign drive sales?

Challenges:

• Unmeasured confounders (hidden factors).
• Selection bias (non-random groups).
• Defining the exact causal mechanism.

Causal inference is crucial in fields like economics, healthcare, and policy-making, where decisions depend on understanding true cause-and-effect relationships.

Matching in Causal Inference

Matching is a statistical technique used to estimate causal effects from observational (non-experimental) data by creating comparable groups—treated and untreated subjects—that are similar on observed characteristics (covariates). This helps mimic randomization, reducing bias from confounding variables.

Key Idea

In an ideal experiment (e.g., a randomized controlled trial), treatment assignment is random, so groups are balanced. In observational data, matching ensures that treated and control units are as similar as possible except for the treatment.

Steps in Matching

1. Define Treatment and Control Groups
   • Treated: Units exposed to the intervention (e.g., drug, policy).
   • Control: Units not exposed.
2. Select Covariates
   • Choose variables that may confound the relationship (e.g., age, income, pre-treatment health).
3. Match Units
   • For each treated unit, find one or more control units with similar covariate values.
4. Assess Balance
   • Check if matched groups are statistically similar on covariates (e.g., using standardized mean differences).
5. Estimate Causal Effect
   • Compare outcomes between matched treated and control units (e.g., average treatment effect on the treated, ATT).

Common Matching Methods

1. Exact Matching
   • Matches units with identical covariate values (rarely feasible in practice).
2. Propensity Score Matching (PSM)
   • Uses a propensity score (probability of treatment given covariates) to match units.
   • Common approaches:
     • Nearest-neighbor matching: Pairs treated units with the closest propensity score in controls.
     • Caliper matching: Only matches units within a specified distance (“caliper”).
     • Stratification: Groups units into strata based on propensity scores.
3. Mahalanobis Distance Matching
   • Matches based on multivariate distance between covariates (accounts for correlations).
4. Coarsened Exact Matching (CEM)
   • Temporarily “coarsens” continuous variables into bins for exact matching.
5. Genetic Matching
   • Uses an algorithm to optimize balance across covariates.

Example

Research Question: Does job training (treatment) increase earnings (outcome)?

• Confounders: Age, education, prior income.
• Matching: For each trained individual, find an untrained person with similar age, education, and prior income.
• Analysis: Compare post-training earnings between matched pairs.

Advantages

• Reduces bias from observed confounders.
• Intuitive—resembles randomized experiments.
• Works well with high-dimensional data when using propensity scores.

Limitations

• No solution for unobserved confounders (e.g., motivation, omitted variables).
• Trade-off between bias and variance:
  • Too strict matching → few matches (high variance).
  • Too loose matching → residual bias.
• Dependence on overlap: Requires treated and control units to share covariate ranges (common support).

Assessing Matching Quality

After matching, check:

1. Balance Statistics (e.g., standardized mean differences < 0.1).
2. Visual Checks (e.g., histograms of propensity scores).