Non-Compliance in Survey Experiments with Placebo Treatments

Abstract

Survey experiments often utilize placebo controls which aim to mimic the treatment conditionwhile leaving the key explanatory variable unaffected. We illustrate the problems that can ariseunder placebo-controlled designs when compliance with treatment assignment is imperfect (e.g.,when respondents are inattentive to experimental manipulations). We argue that estimating com-plier average causal effects (CACEs) under these conditions requires either (i) the inclusion of a“pure control” condition involving no experimental manipulation, or (ii) the strong assumptionthat compliance profiles are balanced across the treatment and placebo arms. We illustrate thesepoints with data from a “failed” survey experiment on how the performance of ethnic minoritiesaffects processes of social categorization.