Working Paper: CEPR ID: DP6157
Authors: Gerard J. van den Berg
Abstract: Instrumental variable estimation requires untestable exclusion restrictions. With policy effects on individual outcomes, there is typically a time interval between the moment the agent realizes that he may be exposed to the policy and the actual exposure or the announcement of the actual treatment status. In such cases there is an incentive for the agent to acquire information on the value of the IV. This leads to violation of the exclusion restriction. We analyze this in a dynamic economic model framework. This provides a foundation of exclusion restrictions in terms of economic behavior. The results are used to describe policy evaluation settings in which instrumental variables are likely or unlikely to make sense. For the latter cases we analyze the asymptotic bias. The exclusion restriction is more likely to be violated if the outcome of interest strongly depends on interactions between the agent's effort before the outcome is realized and the actual treatment status. The bias has the same sign as this interaction effect. Violation does not causally depend on the weakness of the candidate instrument or the size of the average treatment effect. With experiments, violation is more likely if the treatment and control groups are to be of similar size. We also address side-effects. We develop a novel economic interpretation of placebo effects and provide some empirical evidence for the relevance of the analysis.
Keywords: information; placebo effect; policy evaluation; randomization; selection effects; treatment
JEL Codes: C21; C31; C35; C51; D81; D82; D83; D84; J68
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Acquisition of information by agents regarding their treatment status (D82) | Violation of the exclusion restriction (C20) |
| Violation of the exclusion restriction (C20) | Incorrect empirical inference (C59) |
| Interaction between agent's effort and treatment status (C32) | Asymptotic bias of the IV estimator (C26) |
| Strong interaction between treatment and effort (C92) | Significant asymptotic bias of the IV estimator (C26) |
| Similar size of treatment and control groups (C90) | More likely violation of the exclusion restriction (C20) |