We used a staggered difference-in-difference design to assess the relationship between traffic safety interventions and pedestrian and cyclist crash incidence. The units of analysis were site-months. Intersections or street segments (‘sites’) that had a selected intervention installed between 1 January 2015 and 1 December 2023 were eligible for inclusion. Sites where installation occurred between 2015 and 2019 were designated as intervention sites while those with installations occurring between 2020 and 2023 were designated as control sites. The study period included 60 months from 1 January 2015 to 31 December 2019 (ie, the period during which intervention sites were installed and control sites were not yet installed). This procedure supported comparability of the intervention and control sites insofar as NYC identified them as eligible and in need of physical environmental interventions.

We accessed publicly available data for this analysis from NYC Open Data.¹¹ The dependent measure was the presence of any crash that occurred within 100 ft of each site per month. Crash data include all police-reported collisions in NYC in which an individual was injured or killed or in which US$1000 or more worth of damage occurred. Crashes were grouped according to the person injured (cyclist, pedestrian) and severity (injury, fatality).

The independent measure was physical environmental roadway interventions. Three physical intervention categories which each had publicly available data for installations that occurred between 1 January 2015 and 1 December 2023 were chosen: enhanced crossings, speed humps and turn traffic calming (figure 1). Enhanced crossings are marked high-visibility crosswalks in low-vehicle traffic areas near schools, transit centres and parks where there is no traffic signal or stop sign,¹² speed humps are raised areas of roadway that are designed to reduce vehicle speed¹³ and turn traffic calming are a group of treatments designed to encourage safer turns, such as hardened centrelines and slow turn wedges, installed in areas with high frequency of crash injury.¹⁴

For each month and site, we created a binary variable indicating whether crashes had occurred involving cyclist or pedestrian injury or mortality at that time and place. We conducted separate analyses for each intervention type wherein the units of analysis were site-months. There were 169 intervention sites and 16 control sites for enhanced crossings, 1226 intervention sites and 611 control sites for speed humps, and 457 intervention sites and 521 control sites for turn traffic calming. The resultant panel data set for enhanced crossings contained 11 100 site-months (185 sites×60 months); the speed humps data set contained 110 220 site-months (1837 sites×60 months) and the turn traffic calming data set contained 58 680 site-months (978 sites×60 months).

Two-way fixed effects logistic regression models were specified for each exposure (enhanced crossings, speed humps and turn traffic calming) and outcome (cyclist injury or death, pedestrian injury or death) pair. We specified additional models to examine the relationship between turn traffic calming and pedestrian injury and pedestrian fatality separately. All models included fixed effects for both site ID and month to account for unit-specific and time-specific confounding and guard against the small possibility that these interventions affect pedestrian and cyclist volume, and we calculated cluster-robust SEs to account for heteroskedasticity across sites.¹⁵

We assessed the effect modification on the relationship between turn traffic calming and pedestrian outcomes by surrounding street segment length. First, we spatially joined both intervention and control intersections with surrounding street segments and calculated street segment lengths. We identified the length of the longest segment attached to each intersection (a proxy for speeding risk), calculated quartiles for the longest street segment, and assigned these to each intersection. We then conducted stratified analyses to investigate the association between turn traffic calming and crashes involving pedestrian injury or death for each quartile.

We conducted two sensitivity analyses. First, we specified Bayesian conditional autoregressive logistic models, using R-INLA, which calculates the integrated network Laplace approximation of a fully Bayesian model.¹⁶ Models included the independent measure for each intervention, an ‘ever-treated’ variable, covariates for number of adjacent streets and total length of adjacent streets, a fixed effect for month, a random effect for site ID and a conditional autoregressive random effect that controlled for loss of unit independence, identified using a matrix of adjacent intervention and control sites with queen contiguity. Second, we used the did package developed by Callaway and Sant’Anna¹⁷ to estimate the group-time average treatment effect of each intervention of interest, adjusting for the same roadway characteristics as in the R-INLA models. Analyses were performed by using R V.4.3.1 and ArcGIS Pro V.3.0.0.

The public was not involved in the conduct of this study. All data used in this study are publicly available and deidentified.