Covariate matching (nearest-neighbour, Mahalanobis, CEM)

Description

Estimate a treatment effect from observational data by pairing treated units with control units closest on confounders, directly or via coarsened bins, then checking balance.

Where propensity score matching collapses confounders into one score, covariate matching matches on the covariates themselves. Nearest-neighbour matching imputes each treated unit's counterfactual from the M closest controls under a scale-invariant distance (normalised Euclidean or Mahalanobis), with replacement; because matching discrepancies bias the estimate and shrink slowly as covariates multiply, Abadie and Imbens' bias correction adjusts each matched difference using a regression fitted on controls, and variance comes from matching formulas, not the bootstrap. Coarsened exact matching (CEM) instead bins covariates on substantive cutpoints, exactly matches on the bins, drops strata lacking either group and weights the rest; the coarsening bounds imbalance in advance and balance is reported with the multivariate L1 statistic (0 perfect, 1 maximal). Both rest on untestable conditional independence.

Attributes

Loss type
Unmigrated
Stage
Post-launch
Effort
Days

Steps

  1. Choose covariates that satisfy the backdoor criterion from a credible DAG.
  2. Inspect imbalance by comparing means and full distributions of covariates between treated and control groups.
  3. For nearest-neighbour matching, define a distance (normalised Euclidean or Mahalanobis), match with replacement using a small M (often 1 to 4), and inspect matching discrepancies.
  4. Apply bias correction: fit the outcome on covariates among controls and adjust each matched difference by the fitted-value difference.
  5. For CEM, coarsen a copy of the covariates into meaningful bins, form strata, drop strata without both a treated and a control unit, and weight the rest by stratum size.
  6. Estimate the ATT (or ATE) with matching variance formulas or weighted regression on uncoarsened data; do not bootstrap matching estimators.
  7. Report balance in the matched sample (standardised differences or L1 per covariate and overall) and state how pruning changed the estimand.

Sources

  • Causal Inference: The Mixtape (Cunningham, 2021)

Also known as

  • Covariate matching with bias correction
  • Nearest-neighbour covariate matching
  • Mahalanobis matching
  • Abadie-Imbens matching
  • Exact matching
  • Coarsened exact matching
  • CEM

Categories

  • Analytics & Behaviour
  • Experimentation & Causal Inference