Background: The main goal of this project was to quantify the uplift in MTA ridership following the rollout of the congestion pricing program (CPP) in Jan 2025. Vehicles entering the CRZ (local streets and avenues in Manhattan south of and including 60 Street) are charged a $9 toll during peak periods, with the toll dropping to $2.25 outside of peak periods. The expectation is that this will reduce traffic and pollution, as well as increase the use of public transit, the effect of which should be observable through MTA data. The goal of this study was to try to quantify the effect of the uplift on MTA ridership.

Data & Method: I downloaded MTA ridership data from 2022-2024, and trained Prophet models on each of the stations during this time period to make baseline predictions for 2025. Prophet is a reasonable baseline model with user friendly interpretable additive components (trend, seasonality, holidays and other shocks). We assume no growth because the NYC population stayed relatively consistent during the training period, then account for weekly and monthly seasonality, and incorporate NYC-specific holidays. Then we compare the actual ridership levels in 2025, segmented by CRZ vs non-CRZ stations, against the predicted baseline levels in 2025 to determine the amount of uplift attributable to the CPP that went into effect at the beginning of 2025.

This is a one-shot analysis, so the training/test split, model choice, and model parameters were fixed before examining post-policy results. In that sense, this is an unbiased study, as I did not tune the model or parameters until finding an effect (sometimes referred to as p-hacking).

Conclusion: There appears to be a significant uplift in ridership following the implementation of the CPP that was not observable in the baseline data. In addition, there was greater uplift across the MTA stations inside the CRZ compared to outside, which supports the hypothesis that the CPP was effective in the area where the toll was applied.

Limitations: While we can strongly infer that the CPP was effective, an observational study will be limited by the number of variables being controlled for, and can never truly determine causation. By training the prophet model on individual stations, each model learned its own station’s baseline behavior such as seasonality and holiday spikes, but I did not attempt to control for station-level cofounders (e.g. nearby construction projects, road closures, population or demographic changes). However, even if the baseline model is biased, by comparing uplift between CRZ and non-CRZ stations, there is some insulation against systematic bias that isn’t being accounted for in the model parameters - as far as I can tell, it is possible but (hopefully) unlikely that there is a bias that specifically affects CRZ stations differently than non-CRZ stations.