Assignment to Treatment Group on the Basis of a Covariate

Causal Question / Estimand

The average effect on outcome of Treatment 1 versus Treatment 2 over the relevant population, when assignment is governed entirely by an observed covariate . See Causal-Estimand.

Identification Strategy

When the probability of treatment is a known function of alone — assignment is ignorable given — the effect is identified by estimating the conditional expectations in each group and averaging their difference over the distribution of . This is the canonical statement of selection-on-observables (Ignorability) identification, and the conceptual seed of propensity-score methods (assignment as a function of covariates).

Key Assumptions

Ignorability (assignment depends only on the recorded ), Overlap (“substantial overlap in the distribution of ” across groups, or strong prior information), Potential-Outcomes.

Threats to Validity

Sensitivity to the model for the conditional expectations: if the regressions are non-parallel or nonlinear, simple covariance adjustment is biased — Rubin urges searching for nonparallelism/nonlinearity. Coarse blocking degrades estimates. Identification fails if any assignment-relevant variable beyond is omitted.

Setting / Data

n/a — methodological. Running example: assigning children to a compensatory reading program by a reading pretest score .

Key Claims

  • If assignment is based solely on , conditioning on removes confounding — no need to worry about unreliability of or unmeasured background variables.
  • The correct estimator averages the difference in estimated over the population; gain scores and reliability-adjusted scores are generally wrong.
  • Under parallel linear regressions, this reduces to the simple covariance-adjusted estimator.
  • Overlap in (or strong priors) is required for the techniques to be useful.

Connections

Citation

Rubin, D. B. (1977). Assignment to Treatment Group on the Basis of a Covariate. Journal of Educational Statistics, 2(1), 1–26.