System Design · Meta · Medium
Recommended design Use a stratified, business-level cluster-randomized controlled trial with 1:1 allocation. A business-level cRCT is preferable to region-level randomization because there can be only a handful of regions, giving poor power. A stepped-wedge design is worse because time trends and early-pool dynamics are strong confounders. A quasi-experiment should be a fallback only if randomization is operationally impossible. The business is the natural unit of…
Checking your access…