Guillaume Hollard

Séminaires thématiques
big data and econometrics seminar

Guillaume Hollard

Revisiting Randomization with the Cube Method
Co-écrit avec
Laurent Davezies, Pedro Vergara Merino

IBD Salle 21

Îlot Bernard du Bois - Salle 21

5-9 boulevard Maurice Bourdet
13001 Marseille

Mardi 13 février 2024| 14:00 - 15:30

Michel Lubrano : michel.lubrano[at]
Pierre Michel : pierre.michel[at]


In this paper, we introduce a novel randomization procedure for randomized controlled trials (RCTs) designed to improve the utilization of baseline information. We start by offering an overview of prevailing methods employed for unit allocation to treatment. Our investigation reveals a prevalent under-utilization of baseline information by empiricists. Indeed, baseline information is collected before randomization takes places in 90% of RCT published in top-5 journals over the last five years. However, this crucial information is used in only half of the papers incorporate this information for covariate balancing in the randomization process.
The most popular methods (e.g. stratification and pairwise matching) do not ensure perfect balance, especially when dealing with continuous and/or numerous covariates (e.g., stratification and pairwise matching). Other methods, such as rerandomization, limit the scope for robust inference. We here adapt a sampling algorithm, named the cube method the Cube method (Deville and Tillé 2004) and show how it can be used to overcome identifying limitations of existing randomization methods. The Cube method enables the selection of perfectly balanced samples for any covariate, whether continuous or categorical, ensuring consistent adherence to balance tests. Moreover, the method facilitates the generation of unambiguous confidence intervals and yields substantial gains in precision, particularly when covariates exhibit correlations with potential outcomes. The Cube method further allows for the flexible determination of assignment probabilities, permitting variation across subgroups (e.g., to address sample attrition concerns). 
We provide comprehensive theoretical insights and conduct simulation exercises using both randomly generated and real data to illustrate the overarching advantages of the Cube method. Our findings support the contention that the Cube method represents a robust and versatile approach to address the challenges associated with covariate balancing in RCTs, offering researchers an effective tool for experimental design.