Work

At What Level Should One Cluster Standard Errors in Paired Experiments

Year: 2019
Type: preprint
Abstract: In paired experiments, the units included in the randomization, e.g. villages, are matched into pairs, and then one unit of each pair is randomly assigned to treatment. We conducted a survey of papers that used paired randomized experiments in economics. To estimate the treatment effect, researchers usually regress their outcome on a treatment indicator and pair fixed effects, and cluster standard errors at the unit-of-randomization level, namely, at the village level in our example. We show that the variance estimator in this regression may be severely downward biased: under constant treatment effect, its expectation is equal to 1/2 of the true variance. Using that variance estimator may lead researchers to substantially overreject the null of no treatment effect. Instead, we argue that researchers should cluster their standard errors at the pair level.
Language: English
Open Access status: green
OpenAlex ID: W2947628272
Cites:
Cited by: 1
Related to: 10