Mechanism Experiments and Policy Evaluations

Causal Question / Estimand

Argues for a distinct object of study: the effect of a hypothesized causal mechanism, estimated by a “mechanism experiment,” as opposed to the effect of a full policy estimated by a conventional policy evaluation.

Identification Strategy

Both rest on Randomization for internal validity; the contribution is conceptual. A policy evaluation randomizes an actual program and identifies its overall effect. A mechanism experiment instead randomizes a targeted intervention that manipulates a specific channel a policy is presumed to work through — testing the mechanism directly, often more cheaply and before scaling a policy. This sharpens external validity: by isolating a mechanism, results can speak to many policies that share it, addressing the common critique that a single RCT’s LATE-like estimate does not generalize.

Key Assumptions

Randomization, SUTVA. External-validity reasoning assumes the manipulated mechanism is the same one the target policy would operate through.

Threats to Validity

The mechanism studied may not be the one a real policy activates; mechanism experiments trade some policy realism for mechanism clarity; standard RCT threats (noncompliance, spillovers) still apply.

Setting / Data

n/a — methodological/expository (JEP), with policy examples (education, behavioral interventions).

Key Claims

  • Experiments can target mechanisms, not just whole policies, often at lower cost and with broader relevance.
  • Mechanism experiments improve external validity by identifying a channel that many policies share.
  • They complement, rather than replace, policy evaluations in an evidence-based policy pipeline.

Connections

Citation

Ludwig, J., Kling, J. R., & Mullainathan, S. (2011). Mechanism Experiments and Policy Evaluations. Journal of Economic Perspectives, 25(3), 17–38.