Project Restore Synthetic Control Analysis

2025-11-24

Background Information: we are interested in studying the effects of the implementation of Project Restore in Precincts 79/81 (January 2023-January 2024) compared to other precincts in NYC

Issues:

• Precincts 79 and 81 were both exposed to Project Cure in 2015 and 2021, respectively
• Other precincts were also exposed to Project Cure or Project SUV – hard to disentangle effects on violence reductions

Solution:

• We will run 3 analyses:

  1. We will look at precinct 79 as a case alone and use the 11 precincts that implemented Project Cure by 2015 as controls
  2. We will look at precinct 81 as a case alone and use the 6 precincts that implemented Project Cure prior in 2021 as controls
  3. We will look at precincts 79 and 81 together as cases and use the 17 precincts from analyses 1 and 2 as controls.
  • This is probably the best way to keep the control pool comparable for the two cases

Plot of raw data

Initial thought:

• Big decrease in shootings in precinct 79 between 2020 and 2022
• Big decrease in shotoings in precinct 81 between 2022 and 2023

See script 01_analysis_18_24_79_81_combined.R for the recommended analysis (see 01_analysis_18_24_exploratory.R for more detailed analyses for precinct 79)

Recommendation: use ridge, include 2023

Precinct 79

Combined Results

Precinct 81

Combined Results

Precincts 79 and 81 combined

Combined Results

• Findings:

  • Shootings (victims aged 18-24) were reduced by roughly 8.8% (95% CI: -46%, 78.0%) in precinct 79 compared to other similar precincts in NYC, though we note that confidence intervals are wide.
  • Shootings (victims aged 18-24) were reduced by roughly 52.9% (95% CI: -87.0%, 87.7%) in precinct 81 compared to other similar precincts in NYC, though we note that confidence intervals are wide.
  • Shootings (victims aged 18-24) were reduced by roughly 41.1% (95% CI: 5.7%, 57.2%) in precinct 81 compared to other similar precincts in NYC, though we note that confidence intervals are wide.
  • The significance in the combined analysis but not in the individual results may be due to the “smoothed” effect as there is less noise when the trend is averaged out between the two precincts, as well as the increase in the control pool
  • NOTE: we are not adjusting for covariates