Figure 2.

Conditional balance/placebo test. Note: Plots showing conditional balance, given neighborhood and gender. Each pre-treatment covariate is treated as if it were an outcome variable, and the “effect” (imbalance) of the treatment is estimated on each via exact matching on neighborhood and gender. The n shown gives the number of unique treated units plus unique control units contributing to each estimate. Whiskers show the 90% (thick) and 95% (thin) confidence intervals using Abadie–Imbens standard errors.