Every selected subset under the three specifications
Legend: every tag, every symbol and the two scales, defined once
What this table shows, what each column group counts and why the scales are not comparable
What each specification concludes
- Published specification. The published values reproduce: 6 of 6 comparable estimates fall inside our intervals and agree in sign, so the computation is not in question. The significant group is 70.9 per cent Citi Bike stations, individually not significant; its only individually significant component is 14 neckdowns, which are curb reconstructions; no component is the posts or granite blocks the report defines the treatment as.
- Act 2: Extension of the DOT analysis. Held fixed and taken to the study subsets, the corridor DiD design returns 76 significant reductions in 233 estimable cells (87 once the positive controls are included), including cells no mechanism connects, and its own placebo rejects: displaced into months before any installation it still reports significant reductions. A reduction from this design is a property of the comparison group rather than evidence about a treatment.
- Act 3: Method (ours). The placebo test excludes some cells; the counts load with the payload. The single cell surviving everything is an increase in the 2025 sign only cohort on 12 months of follow up, from a design that produces false increases on 78 per cent of null draws under selection on a persistent level. No reduction survives. The defensible summary is that the evidence neither establishes nor excludes a pedestrian injury benefit, and that sign and marking daylighting is a tight null whose own placebo still fails.
The live estimate for this selection
Comparison rule for the live estimate
Subsets that carry no estimate, and why each one carries none
What the live estimate is computed from, and the sums behind the two figures
loading the per unit sums
Run the estimator yourself
The corridor DiD estimator on a window you choose, computed in this page from monthly counts per treated corner. Choose the installation window, the before and after period, the outcome and the comparison rule; the panel reports the estimate, its placebo, and how many units the window admits. The treatments used are the ones selected in the left column.
The monthly payload loads on the first estimate, once, about ten megabytes.
Methodology
The three specifications, their equations and the placebo displacement
Act 1: Replication of the DOT analysis, as published
The design NYC DOT published: the crash record 2017 to 2023, installations 2019 to 2021, 24 months on each side of the installation month, corridor comparison locations on the treated corner's own streets, an unweighted mean of per unit changes in annual injuries, and a two sided one sample t test. The scale is a difference in counts per intersection per year, in which zero means no effect.
For site i, the window count is the outcome inside 30 m of i over the months from a months after installation up to b months after it, and R_i(a,b) expresses that count as an annual rate. The change at the site is the after rate minus the before rate.
LaTeX source
R_i(a,b) \;=\; \frac{12\, Y_i(a,b)}{b-a}LaTeX source
\Delta_i \;=\; R_i(1,\;\text{post}+1) \;-\; R_i(-\text{pre},\;0)The estimate is the unweighted mean over treated sites of the change at the site minus the mean change at that site's own comparison locations, tested against zero on n minus 1 degrees of freedom. Every treated site enters with weight one over n whatever its crash count, so the quantity estimated is the change at the mean treated corner.
LaTeX source
\hat\theta \;=\; \frac{1}{n}\sum_{i=1}^{n}\left(\Delta_i \;-\; \frac{1}{|C_i|}\sum_{j\in C_i}\Delta_j\right)LaTeX source
t \;=\; \frac{\hat\theta}{s_d/\sqrt{n}}Act 2: Extension of the DOT analysis
The same four components, being the corridor comparison rule, the annual rate, the unweighted mean of per unit changes and the one sample t test, carried to the full crash record 2012-07 to 2026-05, to an installation window of eight years running 2015-07 to 2023-06, and to every subset, with 36 months before and up to 24 months after. It reuses the same four equations, so only the crash record, the installation window, the before period and the set of subsets change. The scale is again a difference in counts per intersection per year.
Act 3: Method (ours)
The design of this study: three matched comparison intersections per treated unit, a fixed effects Poisson model carrying a unit effect and a calendar month effect, a cluster robust variance of type CRV1 clustered on the match group, 36 months before and up to 24 months after, a placebo on every cell, and Benjamini and Hochberg adjustment within each outcome. The scale is a rate ratio, in which one means no effect.
LaTeX source
\mathbb{E}\!\left[Y_{kt} \mid \alpha,\gamma,D\right] \;=\; \exp\!\left(\alpha_k + \gamma_t + \beta\, D_{kt}\right)LaTeX source
D_{kt} \;=\; \mathbf{1}\{k \text{ treated}\}\cdot\mathbf{1}\{t > t_{i(k)}\}The month at event time zero is dropped, because an installation month is partly treated. A match group is a treated unit together with its own comparison intersections, and each match group forms one cluster.
LaTeX source
\text{IRR} \;=\; \exp(\hat\beta)Because the specification weights a unit by its expected count, high exposure corners dominate the estimate and the quantity is the proportional change in the total count, while the published estimator gives the change at the mean corner, so the two answer different questions and their magnitudes are not comparable.
The placebo displacement, walked through one case
Number the months of a treated site from its own installation month, so that month zero is the month the device was installed. Take a before period of 24 months and an after period of 24 months, which are the windows of the published specification. A displacement of 24 months runs the placebo window from month minus 48 to month zero, so the last month of the window is the installation month itself, which is partly treated. A displacement of 25 months runs the window from month minus 49 to month minus 1, so the window ends at month minus one and every month of it lies strictly before installation. The displacement is therefore the larger of the before period and one month more than the after period: 25 months in this case, and 36 months on the falsification tested specification, whose before period is 36 months and whose after period is 24.
LaTeX source
s \;=\; \max\!\left(\text{pre},\; \text{post}+1\right)LaTeX source
\left[\,t_i - s - \text{pre},\;\; t_i - s + \text{post}\,\right], \qquad t_i - s + \text{post} \;\le\; t_i - 1The estimate is then recomputed with that one change, holding the corners, the comparison locations, the estimator and the outcome as they were. Because no treatment exists anywhere in the displaced window, the honest answer is no effect, and a cell is read only when its placebo is consistent with no effect at a placebo p of 0.05 or above.
LaTeX source
\hat\theta^{\text{plac}} \;=\; \hat\theta\big(t_i \to t_i - s\big), \qquad \text{admit the cell only if } p^{\text{plac}} \ge 0.05A decrease at p below 0.01 whose placebo returns a decrease of similar size is therefore no evidence about the treatment: the placebo window contains no treatment, so the specification produces that decrease at those corners with nothing installed, and the decrease in the real window cannot be attributed to the installation. The p value answers whether a result could have arisen by chance. The placebo answers whether the specification produces that result without a treatment, and those are different questions.
loading the funnel counts
Adjustment for multiple subsets
Each outcome is estimated over m subsets, being 43 here, so a nominal p value does not control the false positive rate across the family. Benjamini and Hochberg adjustment is applied within each outcome, over the subsets, after sorting the p values in ascending order.
LaTeX source
p^{\text{BH}}_{(k)} \;=\; \min_{j \ge k}\; \min\!\left(1,\; \frac{m\, p_{(j)}}{j}\right)A cell is read only when the adjusted p value falls below 0.05, when the placebo is consistent with no effect, and when at least 25 treated units contribute.
The minimum detectable effect
A null is uninformative without the size of the effect the cell could have found, so every cell carries the minimum detectable effect at 80 per cent power, two sided, which follows from the standard error of the log rate ratio. A null with a bound of 6 per cent excludes a moderate benefit. A null with a bound of 60 per cent excludes almost nothing.
LaTeX source
\text{MDE} \;=\; 100\left(1 - e^{-2.802\,\widehat{\text{se}}}\right)The live estimate for a selection
A browser cannot fit a fixed effects Poisson specification, so a selection is estimated by the published arithmetic estimator, which is exact arithmetic on the per unit window sums, and a descriptive aggregate rate ratio is reported beside it.
LaTeX source
\text{RR}_{\text{agg}} \;=\; \frac{\sum_i T^{\text{after}}_i \big/ \sum_i T^{\text{before}}_i}{\sum_i C^{\text{after}}_i \big/ \sum_i C^{\text{before}}_i}That ratio carries no unit effect, no calendar month effect and no cluster robust variance, so it is not the modelled quantity the manuscript reports and it can disagree in sign with it. Comparison intersections were matched once over the union of treated units, because a browser cannot rematch for an arbitrary selection, so a selection that coincides with a published subset differs slightly from the published estimate for it.
The admission rule and the published rows
Three conditions for reading a cell
The installation date is moved back by the larger of the before period and one month more than the after period. Moving it back only by the before period is not sufficient: with 24 months on each side, that places the final month of the placebo window on t_i itself, which is partly treated. A displacement of at least the after period plus one month ends the placebo window at month minus one, so every month of it is strictly before installation.
A cell is admitted only when three conditions hold together. (i) The placebo is consistent with no effect. (ii) The estimate remains significant after Benjamini and Hochberg adjustment within its outcome, because many subsets are tested on each outcome and a nominal p value does not control the false positive rate over that many tests. (iii) At least 25 treated units contribute, so that a cell resting on a handful of corners is marked underpowered rather than read.
Across every subset and outcome, ten cells are significant after adjustment and all ten have a placebo that rejects no effect, so no cell satisfies all three conditions at once. That is a statement about what this record can support rather than a statement that the treatments do nothing.
Why is the published specification blank for most rows?
The published report contains seven rows: sign only daylighting, hardened daylighting, and the five components of the hardened row. This page reports 40 subsets on 6 outcomes, so for most cells there is nothing published to compare against. The blank is therefore usually a fact about the report rather than about the treatment or about the time range, and the tag says which of four reasons applies.
The report does not form this subset at all. This covers most cells. A subset such as turn calming by quick kurb, or the 2025 published list split by treatment, has no published counterpart because the report never separated it.
The report forms this arm and reports it only on injuries counted as people. It publishes no crash event outcome for any arm, so a row whose outcome counts crashes has nothing to compare against. This covers 12 cells.
The report forms this arm and this project cannot rebuild it, because no record of that treatment carries an installation date inside the published window of 2019 to 2021. This is the case the reader is most likely to guess at, and it applies to Citi Bike stations, whose 5 dated records are dated 2025, and to bicycle corrals, of which 2,833 of 2,858 carry no installation date at all and the remaining 25 are dated 2025.
The report forms this arm and no published inventory of that treatment exists to rebuild it from. This applied to neckdowns and to enhanced crossings. The enhanced crossings have since been located, 197 records of which 195 carry a date, so that arm is now sourceable; neckdowns remain unsourceable, and the agency's own separate evaluation counts 266 of them without publishing any location.
Six cells carry a published value: the three subsets that map onto a published row, each on the two outcomes the report published, being pedestrian injuries and total injuries.
Provenance of the record
The inventory, the Vision Zero View audit and the 266 curb and sidewalk extensions
The inventory does not state what was installed for 907 of 1,491 turn calming records, which are recorded only as a left or right turn treatment, and it carries a single imputed installation month for all 271 records of the January to August 2024 daylighting cohort. An audit of Vision Zero View on 2026-07-29 confirmed that neither gap can be closed from that source: it has no daylighting layer at all across 260 map service layers and 45 catalogue assets, and its turn calming layer carries exactly the four fields already held.
The 266 curb and sidewalk extensions cannot be added to any inventory. No dataset for bulb-outs, curb extensions or neckdowns exists on the open data portal, no such family exists among the Vision Zero View layers, and of 996 street improvement project corridors and 424 intersections only four name a curb extension explicitly.
The audit did add the city's own risk criterion. Treated corners are 3.6 to 20.2 times more likely to be a designated Vision Zero priority intersection than the comparison pool, and 65.5 per cent of hardened daylighting corners lie within 40 m of a street improvement project corridor against 20.7 per cent of the pool. The selection this study previously had to infer from crash trajectories is therefore directly observable.
The hydrant zone idea is the agency's rather than ours. Its January 2025 report contains a hydrant zone analysis, defines a hydrant zone as hydrants and bus stops, and reports a 30 per cent higher normalised injury rate at them over 7,558 such intersections. That analysis is cross sectional; the subgroup test in this study compares treated corners against other treated corners.
The full account, with every figure traced to a file, is in EVIDENCE_PROVENANCE.md in the repository.
Definitions, cautions and provenance
Supplementary panels
The agreement funnel, the calibration, the specification path, the reproduction in full, the component table and the limitations, each one click away.