Heterogeneous Causal Effects of NYC Congestion Pricing

Local Linear Causal Forests | Spatial Policy Evaluation | R

Overview

New York City's Congestion Pricing Act took effect on January 5, 2025, imposing a toll on vehicles entering Manhattan below 60th Street. Early difference-in-differences analysis estimated an average reduction of ~20 motor vehicle collisions per week inside the pricing zone, but average treatment effects mask where and when a pricing intervention actually works.

This project estimates spatially heterogeneous causal effects of congestion pricing on weekly accident rates across NYC's four-borough road network, using Local Linear Causal Forests (LLCF) developed by Friedberg, Tibshirani, Athey & Wager (2021). Rather than a single borough-level average, the model produces a hex-grid map of calibrated conditional average treatment effects (CATEs), one estimate per spatial unit, over a 62-week post-treatment window.

---

Key Findings

• 3,598 hexagonal spatial units covering the majority of NYC were evaluated
• 1 hex reached statistical significance at the 95% confidence level
• Congestion Pricing had no statistically significant impact on vehicle collisions across New York City

---

Methodology

Causal Identification

• Treatment defined as post-January 5, 2025 entry into the congestion pricing zone
• Parallel trends validated across Upper/Lower Manhattan pre-treatment periods
• Unconfoundedness assessed via placebo tests across spatial hex sizes

LLCF Architecture

• Separate nuisance forests (regression_forest) estimate propensity scores (W.hat) and outcome baselines (Y.hat) before the causal forest is trained, the standard R-learner / partially linear setup from Athey & Wager
• Causal forest trained with honesty = TRUE: data split between tree-building and effect-estimation partitions to ensure valid p-values
• Linear correction applied to y_coord and dist_60th at prediction time to sharpen spatial heterogeneity estimates near the cordon boundary
• Standard errors via infinitesimal jackknife variance estimates; significance threshold |t| > 1.645 (90% CL)
• Dynamic calibration via test_calibration(): factor 1.80009 applied to raw CATEs to align predicted effects with observed magnitude
• All forests: 2,000 trees, tune.parameters = "all", clustered by hex_id, parallelized across available cores

Feature Set (X matrix)

Feature	Description
baseline_risk	Pre-treatment mean weekly collisions per hex
y_coord	Northing (EPSG:2263) north-south spatial gradient
dist_60th	Distance from 60th Street cordon boundary
week_index	Continuous time index (weeks since study start)
avg_temp	Weekly average temperature (°F), nearest NYS Mesonet station
tot_precip	Weekly total precipitation (inches), nearest NYS Mesonet station

Data

• NYPD Motor Vehicle Collision Reports (Jan 2022 – Apr 2026); Staten Island excluded
• Spatial grid: 1,640-ft hexagonal tessellation clipped to NYC shoreline via TIGRIS (EPSG:2263)
• Treatment zone: official MTA congestion pricing geofence (WKT); hex assignment by centroid intersection
• Weather: NYS Mesonet monthly CSVs; nearest station per hex via st_nearest_feature; data access fee waived

Computation

• Parallel processing across 64GB RAM local environment
• Post-treatment window: 62 weeks (vs. 10 weeks in prior DiD analysis)

---

Repository Structure

├── ATE_analysis/
│   └── code/
│       └── diff_in_diff             # Original DiD analysis (Upper vs. Lower Manhattan, Co-Author Ellie Stoever)
│
├── HTE_analysis/
│   ├── code/
│   │   ├── Data_Cleaning.R          # Spatial setup, collision panel, weather merge, feature engineering
│   │   └── LLCF_model.R             # Nuisance forests, causal forest, CATE prediction, mapping
│   ├── outputs/
│   │   ├── shapefile                # NYC cordon boundary (MTA geofence)
│   │   ├── Significant_Hexes.png   # Statistically significant CATEs (90% CL)
│   │   ├── All_Hexes.png           # Full spatial distribution (all hexes, no significance filter)
│   │   └── model_output.csv        # Hex-level CATE estimates, SEs, t-stats
│   └── Independent_Study_Introduction.pdf
│
├── Policy_Analysis_SlideDeck.pdf   # In Depth Policy Analysis of the Congestion Pricing Act (2025)
├── Project_Proposal.pdf            # Independent study proposal
├── project_final.pdf               # Final empirical paper (Submission to Advisor by June 15th, 2026)
└── README.md

---

Background and Motivation

This project extends a prior difference-in-differences study (Econ 114, UCSC) that suffered from limited post-treatment data (10 weeks) and weak external validity. Moving to LLCF with 62 weeks of post-treatment observations addresses three limitations of the original analysis:

1. Granularity: Borough-level averages obscure localized policy impacts
2. Heterogeneity: Peak vs. off-peak hours and intersection density drive differential effects that ATE cannot capture
3. Statistical power: A 6x longer post-treatment window yields more reliable causal estimates

The LLCF framework is directly applicable to settings where a single average effect is insufficient, including dynamic pricing, demand estimation, and market design contexts where treatment response varies by unit characteristics.

---

References

• Athey, S. & Wager, S. (2019). Estimating Treatment Effects with Causal Forests: An Application. Observational Studies.
• Tibshirani, J., Athey, S., Sverdrup, E., & Wager, S. (2020). grf: Generalized Random Forests. R package.
• NYPD Motor Vehicle Collisions – Crashes. NYC Open Data.
• NYS Mesonet Weather Data. University at Albany.