Estimating Causal Effects of Treatments in Randomized and Nonrandomized Studies

Causal Question / Estimand

What is the average (“typical”) causal effect of a treatment versus a control, defined as the average of the unit-level differences in Potential-Outcomes ? See Causal-Estimand.

Identification Strategy

The seminal statement of the potential-outcomes framework for causal inference. Rubin defines the causal effect through the two responses a unit would show under treatment and under control, then asks when data can estimate it. Randomization yields unbiased estimates directly; in nonrandomized studies, causal effects can still be estimated by controlling extraneous variation through matching and covariance (regression) adjustment, provided the right assumptions hold.

Key Assumptions

Potential-Outcomes, Randomization (when available), Ignorability (implicit, for the nonrandomized case), Overlap (comparable treated/control units for matching and adjustment).

Threats to Validity

Nonrandomized estimates are biased when assignment depends on unobserved determinants of the outcome (ignorability fails), or when matching/adjustment models are misspecified. Rubin’s stance: prefer randomization, but careful nonrandomized analysis is “a reasonable and necessary procedure in many cases.”

Setting / Data

n/a — methodological. Motivating examples drawn from educational and social-science program evaluation (e.g. compensatory reading programs).

Key Claims

  • Introduces Potential-Outcomes as the basis for defining causal effects, not merely estimating associations.
  • Randomization should be employed whenever possible — but the claim that only randomized experiments yield useful causal estimates is untenable.
  • Matching and covariance adjustment are analysed as devices for estimating causal effects from nonrandomized data.

Connections

Citation

Rubin, D. B. (1974). Estimating Causal Effects of Treatments in Randomized and Nonrandomized Studies. Journal of Educational Psychology, 66(5), 688–701.