2020/12/06 by Abba Μ. Krieger, Krieger, Abba M., David Azriel +3
Mathematics · #Statistical Methods in Clinical Trials #Advanced Causal Inference Techniques #Statistical Methods and Inference
paper · pdf · doi:10.48550/arxiv.2012.03330
We present a new experimental design procedure that divides a set of\nexperimental units into two groups in order to minimize error in estimating an\nadditive treatment effect. One concern is minimizing error at the experimental\ndesign stage is large covariate imbalance between the two groups. Another\nconcern is robustness of design to misspecification in response models. We\naddress both concerns in our proposed design: we first place subjects into\npairs using optimal nonbipartite matching, making our estimator robust to\ncomplicated non-linear response models. Our innovation is to keep the matched\npairs extant, take differences of the covariate values within each matched pair\nand then we use the greedy switching heuristic of Krieger et al. (2019) or\nrerandomization on these differences. This latter step greatly reduce covariate\nimbalance to the rate Op(n-4) in the case of one covariate that are\nuniformly distributed. This rate benefits from the greedy switching heuristic\nwhich is Op(n-3) and the rate of matching which is Op(n-1).\nFurther, our resultant designs are shown to be as random as matching which is\nrobust to unobserved covariates. When compared to previous designs, our\napproach exhibits significant improvement in the mean squared error of the\ntreatment effect estimator when the response model is nonlinear and performs at\nleast as well when it the response model is linear. Our design procedure is\nfound as a method in the open source R package available on CRAN called\nGreedyExperimentalDesign.\n