Robins and Greenland 1996 — Identification of Causal Effects Using Instrumental Variables Comment

Summary

Published as a comment on Angrist, Imbens, and Rubin (AIR, 1996) in the same JASA issue (Vol. 91, No. 434, pp. 456–458), Robins and Greenland argue that the average treatment effect (ATE) in the full study population is often of greater public health interest than the local average treatment effect (LATE) or intention-to-treat (ITT) parameters. They establish that Robins (1989) independently developed the same framework as AIR in the context of AIDS clinical trials, that ATE is not identified under AIR's assumptions alone but can be bounded (Robins-Manski bounds; Balke-Pearl sharp bounds), and that ATE is identifiable under an alternative structural nested mean model (SNMM) assumption that requires neither monotonicity nor constant treatment effects. They also show that in bioequivalence trials (two active treatments), the ITT parameter is often of no public health interest because the sharp bioequivalence null does not imply ITT = 0.

Key Claims

• Robins (1989), analyzing AIDS clinical trial noncompliance, independently assumed the same three conditions as AIR (exclusion restriction, monotonicity, random assignment) and studied ATE bounds and sensitivity analysis — a parallel prior development to AIR's from epidemiology.
• Under AIR's assumptions alone, ATE is not identified — only LATE is. The observed data determines bounds for ATE (Robins-Manski; sharpened by Balke-Pearl 1993). Wide bounds are informative: they reveal how much public health decisions depend on prior beliefs.
• ATE is identifiable without monotonicity under Robins' Assumption 6 — a structural nested mean model (SNMM) implying ATE = IV estimand — provided both (a) within each treatment arm, average treatment effects are equal for treated and untreated; and (b) the average treatment effect among the treated is the same for both arms.
• In bioequivalence trials (new therapy vs. standard), noncompliance means the sharp null of bioequivalence does not imply the ITT null. The ITT test may falsely reject bioequivalence when the result is attributable to differential noncompliance rates, not true drug effects. Neither LATE nor ITT is the right parameter; ATE is, and it requires methods beyond AIR.
• ATE is ITT null consistent (zero iff ITT = 0) under the SNMM, as is LATE. ITT itself is always ITT null consistent.

Concepts Introduced or Extended

• Local Average Treatment Effect — argues ATE (for the full population) is often of greater public health interest than LATE (for unobservable compliers) or ITT; ATE is unidentified under AIR but boundable and identifiable under SNMM
• Instrumental Variables — SNMM (Robins 1989 Assumption 6) identifies ATE as the IV estimand without monotonicity; bioequivalence trials as a setting where LATE breaks down

Entities Mentioned

• James Robins
• Sander Greenland
• Joshua Angrist
• Guido Imbens
• Donald Rubin

Quotes

"We wish to complement the interesting paper by Angrist, Imbens, and Rubin (AIR) by offering several alternative analytic strategies."

"The distribution of the observed data only determines bounds for the ATE parameter."

"The ATE parameter can be of greater public health interest than the LATE or ITT parameter."

My Take

Robins and Greenland are making a genuinely different point from Heckman: not a priority dispute or a critique of AIR's assumptions, but a constructive extension showing that ATE — a richer, policy-relevant parameter — is bounded (not just unidentified), and identifiable under a structural assumption that is actually weaker than monotonicity in some respects. The bioequivalence trial argument is the most practically important: it identifies a real class of settings where the AIR framework (and ITT) are silent. The comment is complementary rather than adversarial, which may explain why the Robins-Greenland tradition and the IV tradition have subsequently converged more than the Heckman-IV divide has.