Measuring Causal Effects When Outcomes Are Hidden

🔎 What This Paper Does

This article develops a clear framework for defining, identifying, and estimating the causal effect of an observed treatment on a latent outcome, labeled the latent treatment effect (LTE). The focus is on latent political concepts such as values, beliefs, and attitudes, and on how experimental or observational treatments can be linked causally to those unobserved traits.

📘 How the LTE Is Defined and Identified

• Introduces the latent treatment effect (LTE) as the causal parameter of interest linking treatment status to an underlying latent outcome.
• Specifies a set of identification assumptions needed to recover the LTE from observed data.
• Highlights an often-overlooked exclusion restriction: treatment must not affect observed indicators except through its effect on the latent outcome.

📊 How the LTE Is Estimated

• Proposes a hierarchical item response model to estimate the LTE directly from item-level indicators of the latent construct.
• Contrasts this one-step hierarchical approach with conventional two-step procedures that first estimate latent scores and then regress them on treatment.

🧪 Evidence from Simulations and Applications

• A simulation study demonstrates that the hierarchical item response estimator yields unbiased LTE estimates when the identification and modeling assumptions hold.
• The same simulations show that conventional two-step approaches tend to produce downward-biased estimates of the LTE.
• The methodology is illustrated using data from two published experimental studies to show practical application.

✨ Why It Matters

This framework and estimator provide a principled way to conduct causal inference when outcomes are latent rather than directly observed. The work underscores the importance of explicitly stating and assessing identification assumptions—especially the exclusion restriction—and choosing estimation strategies that avoid bias introduced by ad hoc two-step procedures.