34.2 Contamination Bias

Goldsmith-Pinkham, Hull, and Kolesár (2022) show regressions with multiple treatments and flexible controls often fail to estimate convex averages of heterogeneous treatment effects, resulting in contamination by non-convex averages of other treatments’ effects.

• 3 estimation methods to avoid this bias and find significant contamination bias in observational studies, with experimental studies showing less due to lower variability in propensity scores.

References

Goldsmith-Pinkham, Paul, Peter Hull, and Michal Kolesár. 2022. “Contamination Bias in Linear Regressions.” National Bureau of Economic Research.