Bounding causal effects in survey experiments with noncompliance or inattention
P1-S14-4
Presented by: Matthew Tyler
Survey experimentalists often want to estimate the effect of a treatment among respondents who actually “receive” treatment. To account for noncompliance and inattention, researchers frequently include post-treatment manipulation or attention checks in their survey experiments. Unfortunately, current methods assume that compliance and attention can be measured without error. In reality, inattentive and noncompliant respondents are particularly prone to measurement errors, leading to biased causal effect estimates. In this paper, I develop a numerical method for sharply bounding causal effects with noisy measures of compliance or attention. To account for sampling variability, I also construct confidence intervals that achieve their nominal asymptotic coverage under a mild differentiability condition. I demonstrate the wide applicability of these methods using recent studies aimed at reducing attitudinal support for political violence.
Keywords: survey experiments, causal inference, partial identification, manipulation checks, attention checks