Main

In 2019, road traffic crashes led to over 1.35 million deaths and 50 million injuries globally1. Of these, about 23% of injuries and 37% of deaths involved pedestrians2. In the US, over 68,000 pedestrians died, and 6.10 million pedestrians were severely injured during 2011–2020 (ref. 3). Large cities like New York City (NYC) bear the majority of this injury and mortality burden, in part due to their large number of residents, greater traffic flows and more regions with high human mobility4.

NYC was among the first adopters of the US Vision Zero program that takes a multidisciplinary, human-centric and systems-based approach to view road traffic crashes as preventable and reduce crash-related morbidity and mortality5. Since 2014, NYC has spent over $850 million on engineering, educational, policy and legislative interventions for road safety, including changes to the physical environment6. Physical environmental interventions are structural changes to roadways and surroundings intended to reduce pre- and peri-event risk of an injury. Examples of such interventions include speed humps, turn-traffic-calming treatments and enhanced crossings, among others. Particularly relevant interventions for crashes involving pedestrians are leading pedestrian intervals (LPIs) that are traffic light sequences providing pedestrians with a brief head start (at least 7 s with a duration of up to 11 s for larger intersections in NYC) to begin crossing an intersection before the turning vehicular traffic is permitted to enter7.

LPIs are one of the most low-cost, easily implementable and common interventions in the US Vision Zero program. However, past evidence on the association of LPIs with reduced crashes and injuries in states across the US and other countries has been mixed8,9,10,11,12,13. Past studies suffer from limitations such as small sample sizes and lack of appropriate comparison groups, among others (Supplementary Box 1).

Intersection-level analysis in a location with a large number of intersections, using a study design accounting for non-random intervention implementation, is required. This study used a rigorous spatial ecological panel design to test the effectiveness of LPIs in NYC in reducing the risk of fatal and non-fatal pedestrian injuries at and around intersections. With 6,003 intersections, NYC provides one of the largest datasets to estimate the effectiveness of LPIs. We used study design, spatial scale and statistical methods that can simultaneously account for multiple threats to internal validity.

Results

From January 2013 to December 2018, about 47.8% of the 6,003 intersections included here had LPIs installed. Supplementary Table 1 provides the distributions of outcomes across treated and control study units (that is, intersection-years). Eighteen percent of the 36,018 intersection-years were treated units. For both treated and control units, 340,541 crashes at and around intersections led to 25,608 non-fatal and 291 fatal pedestrian injuries. LPI-treated intersection-years saw 20.3% fatal injuries and 18.3% non-fatal injuries. Supplementary Fig. 1a–e depicts the outcome distributions across intersections with and without LPIs from January 2013 to December 2018.

The primary analysis using the model adjusted for the regression toward the mean (RTM) found a significant 32.9% (95% confidence interval: 24.7%, 40.2%) reduction in the risk of total injuries at LPI-treated units compared with those without LPIs (Fig. 1). With a baseline risk of 42.5% among the control units, the corresponding absolute risk reduction was 6.6 total injuries per 100 intersection-years (Supplementary Table 2). A significant reduction of 32.8% (24.6%, 40.1%) was also observed for non-fatal injuries. With a baseline risk of 42.2%, the corresponding absolute risk reduction was 6.6 non-fatal injuries per 100 intersection-years. The reductions were greater in RTM-adjusted models compared with the crude models (Fig. 1). Sensitivity analyses using year-lagged treatment and turn-traffic-calming-adjusted models confirmed the findings of the RTM-adjusted models for total and non-fatal injury risk (Supplementary Fig. 2). LPIs did not significantly reduce the risk of fatal injuries (48% [−3.2%, 73.8%]), though the effect size was large (Fig. 1). Furthermore, the point estimates for the fatal injury risk were on different sides of the null value (adjusted odds ratio, aOR = 1) for RTM-adjusted and crude models (aOR = 0.5 versus 1.1, respectively). The sensitivity analyses agreed with the direction of effect of the RTM-adjusted model, with year-lagged treatment model showing a significant reduction of 81.4% (34.8%, 94.7%) in fatal injury risk (Supplementary Fig. 2).

Fig. 1: Forest plot for effectiveness (aOR values) of LPI on total, non-fatal and fatal pedestrian injuries.
figure 1

Both models had fixed effects for years and random effects for intersections. The models use 6,003 intersections resulting in 36,018 study units or observations (that is, intersection-years). The adjusted model refers to adjustment for RTM. Raw numbers are in the accompanying text and the chart shows log-transformed aOR values with accompanying 95% confidence intervals. The null value is at 0 in the chart.

Source data

Full size image

For subgroups based on the time of the day, we found that LPIs significantly reduced the risk of total, non-fatal and fatal injuries during daytime and total and non-fatal injuries during nighttime (Supplementary Table 3).

After testing the variables hypothesized to produce different strengths of effect, we did not find significant effect measure modification by the number of intersection legs, sum of segments widths, that is, proxy for intersection perimeter, or length of the longest leg entering the intersection for the risk of total pedestrian injuries (Table 1).

Table 1 Effect measure modification with aOR values for the interaction terms
Full size table

Discussion

LPIs reduced the total pedestrian injuries at and around intersections in NYC by one-third. The sensitivity, effect measure modification and subgroups analyses confirmed the findings’ robustness and generalizability across intersections in NYC. These findings support the use of LPI in NYC and other cities as an effective intervention to improve pedestrian safety.

Adjusting for the RTM improved the risk reductions across pedestrian injury outcomes, with greater improvements noted for fatal injuries. As noted in Supplementary Methods, this supports our rationale that such correction is important to understand the intervention effects in the spatial ecological panel design in which site selection is based on the presence of outcome in the previous years. The sensitivity analyses noted that the effectiveness of LPI across outcomes was not altered in any significant way by the presence of turn-traffic-calming measures. However, considering a year-lagged treatment only improved the effect sizes of the RTM-adjusted models, especially for fatal pedestrian injuries.

Subgroup analyses noted greater pedestrian injury risk reductions during daytime than that for nighttime, potentially due to better visibility during the daytime. The effect modification analyses showed that LPIs are effective regardless of the street-level characteristics of the roadway segments.

Although the risk reductions in fatal injuries were large in magnitude, the confidence intervals included the possibility of no association. There are three possible explanations for this finding. First, the strong association could be due to random statistical chance, though we consider this possibility to be unlikely given the strong analogous associations for non-fatal injuries, the compelling theory connecting the intervention to the outcome and the robust findings in the sensitivity analyses. Second, LPIs are designed to reduce conflicts between turning motor vehicles and pedestrians, but fatalities are more common when straight-traveling motor vehicles go faster and lead to a high-impact crash. That phenomenon would attenuate the associations for fatal compared with non-fatal crashes, so also consider that possibility to be unlikely. Finally, and most likely, pedestrian fatal injuries were rare, and the sparse outcome produced wide standard errors that exceeded the statistical power provided by the available sample.

Results of this study accord with findings from prior evaluations on the association of LPIs with reduced pedestrian crash risk. Previous estimates for LPIs have varied from 12% to 95% reduction in pedestrian-involved crashes across different cities, outcome measures and analytical methods10. However, the majority of studies relied on a small number of intersections (n = 3–105). Specifically, for NYC, three past studies conducted at different times from the mid-1990s to mid-2010s noted 28–37% reductions using 26–104 intersections11,12,13. In contrast, we used one of the largest samples of intersections (n = 6,003), allowing us to produce more precise effect-size estimates. Previous studies used pre–post or case-control designs11,12,13. Our rigorous design ensured better comparability between the treated and control units and the model specification allowed for random effects across intersections. Furthermore, using intersection as the geographic unit of analysis is in line with the mechanism of action of LPIs. That said, LPIs are effective in the context of various environmental, educational, enforcement and legislative interventions under the Vision Zero program that aims for a holistic change to achieve road traffic safety.

The current study has several limitations. First, the temporal unit of analysis (years) may have led to the misclassification of outcome occurrence relative to the time of treatment. However, our year-lagged treatment sensitivity analysis showed the robustness of effect for total and non-fatal injuries. Furthermore, selecting years helped obviate seasonality and temporal autocorrelation in outcomes. Second, spatial autocorrelation may have biased the estimates. However, we used spatial buffers to include only the crashes within 100 feet (30.48 m) of the intersections, which ensured the independence of intersections from one another. Third, confounding due to effects of other physical, awareness and policy-level interventions, as well as variables such as pedestrian infrastructure, traffic volume, turning versus straight-moving vehicles and so on, is plausible. Available data only allowed us to adjust for turn-traffic-calming measures, which did not show any major changes in the effect of LPI on injury outcomes. However, prospective data collection in the future should consider more interventions and roadway and traffic variables. Fourth, our approach using spatial joins could not confirm if the LPIs were present on all arms of an intersection. However, our spatial units were intersections; hence, differences between arms at an intersection were not a focus. Any such within-intersection differences should be subsumed under our models’ random-effects assumption. Future studies could use street view information for confirming arm-specific LPI installations and their functional status. Finally, the findings may not be transportable to other settings including smaller cities, given the uniqueness of the NYC context. For example, NYC LPI durations are longer than those in other US cities (typically 3–7 s). However, we estimated effects at the intersection level and also determined that these are consistent with the assessed effect measure modifiers, marking the generalizability to other parts of NYC and cities with similar pedestrian and vehicular traffic volumes.

Conclusion

We present a novel intersection-level assessment of NYC, demonstrating that LPIs are effective in reducing the risk of pedestrian injuries at and around intersections. Our analysis supports implementing LPIs at about 30,000 currently untreated intersections in NYC. LPIs might also be an important intervention to consider for planners and policymakers of cities in which increased risk for pedestrian-involving crashes at intersections is a major public health concern. Future work should test whether bundling LPIs with other physical environmental interventions such as turn-traffic calming, enhanced crossing and so on at and around intersections has synergistic protective effects.

Methods

The study setting was NYC from January 2013 to December 2018. We used a spatial ecological panel design to assess the risk reduction in pedestrian injuries, with intersection-year as the unit of analysis (Supplementary Methods). NYC Open Data and NYC Vision Zero websites provided precise geolocations (latitude and longitude coordinates) of the outcome, exposure/intervention and other variables used as covariates and effect measure modifiers14.

Variables

Outcomes

We considered non-fatal and fatal pedestrian injuries in collisions with motor vehicles within 100 feet (30.48 m) of a signalized intersection in a year as outcomes. We conducted spatial joins between crash and intersection files to filter the eligible crashes. We chose a 100-feet buffer as it approximates the size of an intersection, and it is probably causally associated with the intersection traffic flow.

Treatment

We considered signalized intersections that are within 10 feet (3.05 m) of LPIs as treated, whereas those beyond the 10-feet buffer of the LPIs were considered untreated. Installment year denoted treatment initiation. Hence, the treatment status was dependent on both location and time.

We included signalized intersections that had LPIs installed during January 2013 and December 2023. We defined intersections in which LPI installation occurred from January 2013 to December 2018 as the treated sites. We designated those with installations occurring after December 2018 as the control sites. Our approach provided comparability between treated and control (not yet treated) sites insofar as NYC noted them as eligible and needing LPIs.

Effect measure modifiers

We considered three modifiers: number of roadway segments meeting at the intersection, also known as legs (ordinal coding of 3, 4 and >4 leg intersections), sum of segment widths that proxies the intersection perimeter (coded as quartiles), and length of the longest roadway segment or leg entering the intersection, measured from the next intersection along that roadway to the study intersection itself (coded as quartiles).

Subgroup categories

We grouped crash events by the time of the day. Two twelve-hour categories for the time of the day included daytime (7:00 a.m. to 6:59 p.m.) and nighttime (7:00 p.m. to 6:59 a.m.).

For more details on the modifier and subgroup variables, see the Supplementary Methods.

Analysis

We used mixed-effects logistic regression to estimate the treatment effect of LPIs in separate models for the odds of total, non-fatal and fatal injuries. In the main analysis, along with the treatment variable of interest, all the models included fixed effects for years, random effects (intercepts) for intersections, a term for the “ever-treated” status of the intersection and a preintervention outcome mean term to adjust for the RTM (Supplementary Methods provides the rationale). We also analyzed crude models without the RTM adjustment term.

The aOR values approximate the relative risk since the outcome count at each intersection was relatively rare. Robust standard errors account for heteroskedasticity across intersection sites. We estimated absolute risk reduction per intersection-year using the aOR values and baseline risk estimates, that is, estimates of the group-wise marginal mean probability.

We conducted two sensitivity analyses. First, to test for temporal mis-specification of the outcome events relative to LPI installment using a year-lagged treatment variable. For example, for an intersection with LPI installed in 2014, we considered the treatment effect for 2015 and beyond. Second, to test if the presence of another intervention at the intersection confounded the effect of LPI, we adjusted for the presence of turn-traffic-calming measures (Supplementary Methods).

We conducted effect measure modification analysis using similar mixed-effects models for total injury risk as the outcome and interpreted the aOR values of the interaction terms to investigate significant effect modification. The ordinal coding ensured model parsimony. Use of quartiles ensured that the outliers in the distribution of a modifier did not skew the effect.

We conducted subgroup analyses for total, non-fatal and fatal injury risk outcomes based on two categories of time of the day. Hence, we ran two mixed-effects models for each outcome. The models for sensitivity, effect measure modification and subgroup analyses were RTM adjusted.

We used ArcGIS Pro v. 3.0.0 and QGIS v. 3.36.0 for geospatial data preprocessing; Stata v. 18.0 for statistical modeling; Datawrapper for visualization; R v. 4.4.3 for data wrangling, modeling and visualization; and Microsoft Excel 16.29 for validation.

Ethics statement

We used publicly available event-level data with no personally identifiable information. Hence, the study did not need an IRB review or approval.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.