Marginal Treatment Effect

Definition

The Marginal Treatment Effect (MTE) is the average treatment effect for individuals who are at the margin of indifference between treatment and non-treatment at a specific value of the latent selection index (or equivalently, the propensity score). $\text{MTE}(u) = E[Y_1 - Y_0 \mid V = u]$, where $V$ is the unobservable that drives selection into treatment and $u$ indexes where in the propensity score distribution the marginal individual lies. Individuals with $V < p(Z)$ are induced into treatment by instrument values that generate propensity score $p(Z)$; MTE at threshold $u$ represents the average treatment effect for the subpopulation exactly at that margin.

Key Ideas

• Origin: Introduced by Björklund and Moffitt (1987) to study self-selection into training programs within a Roy-model framework. Developed into a full unifying framework by Heckman and Vytlacil (1999, 2005).
• Unifying estimator weights: Heckman and Vytlacil show that all standard estimators identify different weighted averages of MTE, with weights determined by the instrument:
  • Ordinary Least Squares (OLS): weighted average of MTE with weights reflecting the density of the propensity score at each threshold, upward-biased when individuals with high private returns select into treatment (essential heterogeneity).
  • Instrumental Variables (IV)/Local Average Treatment Effect (LATE): weighted average of MTE at propensity-score thresholds where the instrument has identifying variation — the weights depend on the instrument's distribution and first-stage strength, which is why different instruments yield different LATEs.
  • Matching/average treatment effect on the treated (ATT): weighted average of MTE integrated over the treated population's propensity score distribution.
  • difference-in-differences (DiD): weighted average of MTE for the subset of individuals whose treatment status changed between periods due to the policy change.
• Essential heterogeneity: When individuals select into treatment partly because they privately know their own above-average return ($E[Y_1 - Y_0 \mid V \leq u]$ increases as $u$ decreases — high-private-return people sort into treatment first), OLS overestimates average treatment effect (ATE) and IV estimates are sensitive to which complier margin the instrument identifies. MTE makes this selection-on-gains explicit.
• MTE and the propensity score: MTE is identified by the derivative of $E[Y \mid X, Z]$ with respect to $P(Z)$ at a given propensity score value, under local IV conditions. With a continuous instrument, MTE is traced out across the full unit interval; with a discrete instrument, MTE is identified at a finite set of points.
• Policy-relevant treatment effects: For any well-specified policy change, the relevant treatment effect is a particular weighted average of MTE with weights determined by the marginal distribution of the propensity score before vs. after the policy. This generalizes LATE by allowing the policy-relevant weights to be calculated from the structural model rather than forced to equal the instrument-specific LATE weights.

Why It Matters

The MTE framework resolves the debate between the econometric structural approach (Heckman) and the reduced-form IV approach (Angrist-Imbens) by showing they estimate the same underlying object — MTE — with different weights. The disagreement is therefore not about whether to use IV or structural models, but about which weighted average of MTE is policy-relevant. For P1 questions (evaluating a past program for the treated), LATE may suffice if the instrument mimics the policy. For P2 and P3 questions (new environments, new policies), recovering the full MTE curve and re-weighting with new policy weights requires structural assumptions. MTE also diagnoses essential heterogeneity: a declining MTE curve (higher returns at lower propensity-score thresholds, i.e., for more reluctant participants) signals that current treated individuals have higher returns than the marginal participants who would be induced by a policy expansion.

Open Questions

• Identifying the full MTE curve requires a continuous instrument with sufficient support — rare in practice. What are valid approximations with discrete instruments?
• When MTE varies with observed covariates $X$, estimating the heterogeneous MTE surface is high-dimensional. What regularization or aggregation strategies are valid?
• Can MTE estimates from different studies or instruments be combined to trace out a common MTE curve?

Related

• Heterogeneous Treatment Effects
• Instrumental Variables
• Local Average Treatment Effect
• Rubin Causal Model
• Natural Experiments
• Difference-in-Differences
• James Heckman
• Joshua Angrist
• Guido Imbens
• Jennie Brand

Sources

• Heckman 2008 — Econometric Causality
• Heckman 1996 — Identification of Causal Effects Using Instrumental Variables Comment
• Cunha and Heckman 2007 — Identifying and Estimating the Distributions of Ex Post and Ex Ante Returns to Schooling
• Brand and Simon-Thomas 2012 — Causal Effect Heterogeneity
• Brand and Xie 2010 — Who Benefits Most from College