Working Paper: NBER ID: w17408
Authors: Alan I. Barreca; Jason M. Lindo; Glen R. Waddell
Abstract: This study uses Monte Carlo simulations to demonstrate that regression-discontinuity designs arrive at biased estimates when attributes related to outcomes predict heaping in the running variable. After showing that our usual diagnostics are poorly suited to identifying this type of problem, we provide alternatives. We also demonstrate how the magnitude and direction of the bias varies with bandwidth choice and the location of the data heaps relative to the treatment threshold. Finally, we discuss approaches to correcting for this type of problem before considering these issues in several non-simulated environments.
Keywords: Regression Discontinuity; Heaping; Causal Inference
JEL Codes: C14; C21; I12
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Dropping observations at data heaps (C55) | Unbiased estimates for continuous types (C46) |
| Nonrandom heaping in regression discontinuity designs (C22) | Regression discontinuity estimates are biased (C22) |
| Choice of bandwidth (C46) | Magnitude and direction of bias (C46) |
| Heaps not located near the treatment threshold (L99) | Nonrandom heaping can introduce bias (C83) |