Propensity Score Matching

The core function of propensity score matching is to reduce selection bias by creating a balanced comparison group. It does this by matching each treated unit with one or more control units that have a similar propensity score—the estimated probability of receiving the treatment given their observed covariates (e.g., age, income, prior behavior).

• Goal: Make the treatment and control groups statistically comparable on observed variables, mimicking randomization.
• Result: Differences in outcomes between the matched groups can be more credibly attributed to the treatment effect, not pre-existing differences.

The propensity score, denoted as e(X), is a single scalar summary of all observed pre-treatment covariates (X). It is defined as the conditional probability of assignment to a treatment given the observed covariates: e(X) = Pr(T=1 | X).

• Estimation: Typically estimated using a logistic regression or a machine learning classifier (e.g., gradient boosting) where the treatment assignment is the dependent variable and covariates are predictors.
• Role: According to the Rosenbaum-Rubin Theorem, if treatment assignment is strongly ignorable given X, then it is also strongly ignorable given the propensity score e(X). This allows for matching on a single dimension.

Once propensity scores are estimated, units are paired using specific algorithms. The choice affects the quality and variance of the causal estimate.

• Nearest Neighbor Matching: Each treated unit is matched with the control unit whose propensity score is closest. Can be performed with or without replacement.
• Caliper Matching: A tolerance level (caliper) is set (e.g., 0.2 standard deviations of the propensity score). Matches are only made if the score difference is within this caliper, improving match quality.
• Stratification/Subclassification: Units are divided into strata (e.g., quintiles) based on their propensity score. The treatment effect is estimated within each stratum and then averaged.
• Optimal Matching: Minimizes the total absolute distance across all matches, often producing more balanced samples than greedy nearest-neighbor.

Propensity score matching relies on the critical Strong Ignorability or Unconfoundedness assumption. This has two parts:

1. Conditional Independence: The potential outcomes (Y(1), Y(0)) are independent of the treatment assignment (T) given the observed covariates X. Formally: (Y(1), Y(0)) ⟂ T | X.
2. Positivity/Overlap: For all possible values of X, there is a positive probability of receiving either treatment or control. Formally: 0 < Pr(T=1 | X) < 1.

• Implication: This assumption means there are no unobserved confounders. Violation of this assumption (i.e., hidden bias) invalidates the causal conclusions from PSM.

After matching, analysts must check if balance was achieved. This is a crucial validation step.

• Standardized Mean Difference (SMD): The primary metric. For each covariate, calculate the difference in means between treated and control groups, divided by the pooled standard deviation. An SMD below 0.1 is typically considered good balance.
• Variance Ratios: The ratio of variances for each covariate between groups should be close to 1.
• Visual Checks: Examine plots like love plots (forest plots of SMDs before/after matching) and propensity score distribution histograms to assess overlap improvement.

PSM is a quasi-experimental method used when randomized controlled trials (RCTs) are infeasible, unethical, or too costly.

• RCT Gold Standard: Random assignment ensures groups are balanced on both observed and unobserved covariates on average.
• PSM Limitation: Only balances observed covariates. It cannot adjust for unobserved confounders, which remains its fundamental weakness.
• Use Case: Commonly applied in observational studies in economics (e.g., evaluating job training programs), healthcare (e.g., drug effectiveness from electronic health records), and marketing (e.g., measuring campaign impact from customer data).

Causal inference is the process of drawing conclusions about cause-and-effect relationships from data. Unlike correlation, it seeks to estimate the impact of an intervention (treatment). Core methods include:

• Randomized Controlled Trials: The gold standard, where random assignment eliminates confounding.
• Quasi-Experimental Methods: Used when randomization isn't possible (e.g., propensity score matching, difference-in-differences).
• Structural Causal Models: A framework using directed acyclic graphs to encode assumptions about data-generating processes.

The Average Treatment Effect is the primary target of causal inference. It represents the average difference in outcomes between a treatment group and a control group across a population.

• ATE: The effect for the entire population.
• ATT: Average Treatment Effect on the Treated (the effect for those who actually received the treatment).
• ATC: Average Treatment Effect on the Controls. Propensity score matching is often used to estimate the ATT, answering: 'What was the effect for those who received the treatment?'

Selection bias is the systematic error that occurs when the treated and untreated groups differ in ways that affect the outcome, independent of the treatment itself. This is the fundamental problem propensity score matching aims to solve.

• Sources: Self-selection, non-random program enrollment, or confounding variables.
• Consequence: Observed correlation does not equal causation. For example, comparing outcomes of patients who chose a drug versus those who didn't may reflect underlying health differences, not just drug efficacy.

A confounding variable is a factor that influences both the treatment assignment and the outcome, creating a spurious association. It is the primary driver of selection bias.

• Example: In studying a training program's effect on salary, prior education confounds the analysis if more educated individuals are both more likely to take the program and earn higher salaries regardless.
• Role in PSM: Propensity score matching attempts to balance observed confounders (e.g., age, education, income) between the treated and matched control units, simulating randomization.

Stratified sampling is a probability sampling technique where a population is divided into homogeneous subgroups (strata) based on key characteristics before sampling. It ensures all subgroups are adequately represented.

• Relation to PSM: Propensity score matching can be viewed as a form of post-hoc stratification. Instead of pre-defining strata, units are grouped by their estimated propensity score (e.g., 0.0-0.1, 0.1-0.2), and matching occurs within these 'score strata' to improve balance.

Instrumental Variables is an alternative causal inference method used when unobserved confounding is suspected. It relies on finding a variable (the instrument) that:

1. Correlates with the treatment assignment.
2. Affects the outcome only through its effect on the treatment (exclusion restriction).

• Comparison to PSM: While PSM addresses observed confounding, IV methods aim to handle unobserved confounding. However, finding a valid instrument is often challenging. The methods are complementary tools in the causal inference toolkit.