Brand and Simon-Thomas 2012 — Causal Effect Heterogeneity

Summary

Brand and Simon-Thomas review the methodological landscape for studying causal effect heterogeneity — variation in treatment effects across individuals — with particular attention to propensity-based approaches. The chapter argues that standard regression analysis gives misleading representations of causal effects when effects vary across individuals, and encourages social scientists to routinely examine how effects differ across the propensity-score distribution as well as across observable subgroups. The central empirical example is the effect of college attendance on civic participation. Published in Morgan (ed.), Handbook of Causal Analysis for Social Research (Springer, 2013); this draft circulated in 2012.

Key Claims

• Standard regression coefficients implicitly assume homogeneous treatment effects; when effects actually vary across individuals, a single estimated coefficient misrepresents the distribution of effects and can mislead policy recommendations.
• Researchers should routinely characterize heterogeneity both by observable covariates (subgroup analysis) and by treatment propensity — estimating conditional average treatment effects within propensity strata reveals patterns invisible to covariate-based analysis alone.
• Negative selection: In some settings, individuals least likely to receive a treatment benefit most from it. This is the empirical finding of Brand and Xie (2010) for college wage returns — those with the lowest propensity to attend college show the largest earnings gains from attending. This directly contradicts the standard assumption of positive selection (where the most able, who are most likely to attend, also benefit most).
• The four standard estimands — average treatment effect (ATE), average treatment effect on the treated (ATT), local average treatment effect (LATE), marginal treatment effect (MTE) — each captures heterogeneity at a different margin. LATE is instrument-specific and identifies effects for compliers only; ATE averages over the full population; ATT averages over the treated; MTE traces effects as a continuous function of the propensity score threshold.
• Propensity-based stratification (estimating treatment effects within propensity score strata) provides a practical nonparametric approach to heterogeneity that is more transparent than parametric interaction models and more interpretable than MTE curves.

Concepts Introduced or Extended

• Heterogeneous Treatment Effects
• Marginal Treatment Effect
• Local Average Treatment Effect

Entities Mentioned

• Jennie Brand
• Juli Simon-Thomas
• Stephen Morgan

Quotes

"Regression analysis can offer misleading representations of causal effects that vary across individuals... researchers should examine treatment-effect heterogeneity with the same rigor they devote to pretreatment heterogeneity."

"Individuals who are least likely to receive a treatment, due to background characteristics and circumstances, may be the ones who benefit from it most."

My Take

A useful review chapter that makes the case for routine heterogeneity analysis in a social science audience not always fluent in the MTE/econometrics tradition. The positive/negative selection distinction is the conceptually sharpest contribution — it reframes the external validity question ("does this LATE generalize?") in terms of how treatment effects vary with selection propensity. The civic participation empirical example is pedagogically clear but thinner than the wage returns application in Brand and Xie (2010). The chapter's main limitation is that propensity-score stratification methods rely on selection-on-observables; for settings with unobservable selection on gains, the MTE framework (Heckman and Vytlacil) is required, and this chapter does not fully integrate the two traditions.