Working Paper: NBER ID: w30745
Authors: Sylvain Chassang; Rong Feng
Abstract: Clinical trials following the “gold standard” of random assignment frequently use independent lotteries to allocate patients to treatment and control arms. However, independent assignment can generate treatment and control arms that are unbalanced (i.e. treatment and control populations with significantly different demographics), which reduces power. Other assignment methods such as matched pair designs ensure balance across arms while maintaining randomization and permitting inference. \nThis paper seeks to measure the cost of imbalance with respect to gender in a sample of roughly 2000 clinical studies. We document significant imbalance: 25% of experiments have at least 26% more men in one treatment arm than in the other. In addition, clinical trials with greater imbalance have more dispersed treatment effects, indicating that imbalance reduces the informativeness of experiments. A simple structural model suggests that for a typical experiment, using a balanced random design could deliver informativeness gains equivalent to increasing the sample size by 18%.
Keywords: No keywords provided
JEL Codes: C90; C93; I18; I19
Edges that are evidenced by causal inference methods are in orange, and the rest are in light blue.
| Cause | Effect |
|---|---|
| Imbalance in gender across treatment arms (J16) | less informative experiments (C90) |
| Imbalance (D59) | clarity of treatment efficacy (C22) |
| Balanced random design (C90) | informativeness gains (D83) |
| Absolute imbalance (D59) | treatment effect size (C21) |
| Sample size control (C90) | omitted variable bias (C20) |