Professional Documents
Culture Documents
A R T I C L E I N F O A B S T R A C T
Keywords: Assessment of collision risk between pedestrians and automobiles offers a powerful and informative tool in urban
Pedestrians planning applications, and can be leveraged to inform proper placement of improvements and treatment projects
Safety to improve pedestrian safety. Such assessment can be performed using existing datasets of crashes, pedestrian
Collisions counts, and automobile traffic flows to identify intersections or corridors characterized by elevated collision risks
Urban planning
to pedestrians. The Safety In Numbers phenomenon, which refers to the observable effect that pedestrian safety
is positively correlated with increased pedestrian traffic in a given area (i.e. that the individual per-pedestrian
risk of a collision decreases with additional pedestrians), is a readily observed phenomenon that has been studied
previously, though its directional causality is not yet known. A sample of 488 intersections in Minneapolis were
analyzed, and statistically-significant log-linear relationships between pedestrian traffic flows and the per-pe-
destrian crash risk were found, indicating the Safety In Numbers effect. Potential planning applications of this
analysis framework towards improving pedestrian safety in urban environments are discussed.
⁎
Corresponding author.
E-mail addresses: murph677@umn.edu (B. Murphy), dlevinson@umn.edu (D.M. Levinson), aowen@umn.edu (A. Owen).
http://dx.doi.org/10.1016/j.aap.2017.06.004
Received 12 August 2015; Received in revised form 16 May 2017; Accepted 6 June 2017
Available online 15 June 2017
0001-4575/ © 2017 Elsevier Ltd. All rights reserved.
B. Murphy et al. Accident Analysis and Prevention 106 (2017) 181–190
vehicles decreases as a function of the increasing flow of pedestrian morning peak (7–9 AM), midday (11 AM–1 PM), and evening peak
traffic. SIN is well-supported by pedestrian crash data across a number (4–6 PM); these were summed together to produce a 6-h count, re-
of studies in various urban environments and reviews (Jacobsen, 2003; presenting pedestrian traffic when a higher number of cars are on the
Leden, 2002; Bhatia and Wier, 2011). The concept has seen relatively road. The counts took place sporadically across the 14-year window,
widespread adoption in urban planning schools of thought, though its and most intersections were only counted once or twice due to the
temporal causality is not clear-cut (Bhatia and Wier, 2011), and it is rotating schedule on which counts occurred. An “annual average daily
commonly discussed only in the context of pedestrian risk depending on 6-h count” was obtained by averaging the intersection-counts over the
pedestrian flow levels. The United States Department of Transportation number of years for which that intersection was counted. AADT mea-
(USDOT) Strategic Plan for Fiscal Years 2012–2016 aims to reduce non- surements were associated with street links, not to intersections. To
vehicle-occupant fatalities to 0.15 per 100 million vehicle-miles-tra- associate AADT numbers with intersections, the AADT for each unique
veled (VMT) by 2016. However, such a goal does not account for risk street at an intersection (typically two) was summed together. Crashes
dependence on pedestrian flow levels, and thus the federal guidelines were tabulated to include both non-fatal and fatal crashes. Finally,
ignore the SIN effect. manual geocoding of the three datasets (TMC, AADT, and crash counts)
By necessity, data informing placement of improvements and pro- to the intersection spatial layer took place, to allow spatial analysis.
jects for walking and bicycling safety, such as pedestrian bump-outs At a given intersection, “pedestrian risk” is defined as the number of
(Wolshon and Wahl, 1999) and traffic-calming measures (Campbell crashes, between cars and pedestrians, that occurred across the 14-year
et al., 2004), must be sufficiently granular; travel behavior studies are period, divided by the 6-h count of pedestrians; this gives a metric to
typically performed at the Transportation Analysis Zone (TAZ) level, assess the risk of being hit by a car that an individual pedestrian may
which is insufficiently fine-grained to allow for analysis of the shorter- experience at such an intersection. “Car risk” is defined as the number
distance travel modes of bicycling and walking. Schneider et al. (2004) of crashes that occurred across the analysis period, divided by AADT,
provides a relevant framework for building a comprehensive pedestrian which gives a metric to assess the risk of being involved in a car-pe-
risk assessment model, with a granular focus on a specific university destrian collision which a driver may experience. How these risks vary
campus and a model which included factors of pedestrian flow, vehicle from intersections with low to high traffic flows determines whether
flow, and an environmental factor (crosswalk length). Wier et al. (2009) pedestrians or cars experience SIN.
provide precedent for area-level modeling of pedestrian risk in- The sample of intersections to use within the analysis was created
corporating zoning and land use characteristics. These levels of detail by identifying intersections with both nonzero pedestrian counts and
correlate well with the realities of implementation of pedestrian safety nonzero AADT data; there were 448 such intersections, and a summary
investments, which occur not on the city-wide level, but within specific of the sample data can be found in Table 2 in Section 3. A map of the
intersections and road segments. intersections included in the analysis is shown in Fig. 1. Of note, the 6-h
Pedestrian traffic counts, Average Annual Daily Traffic (AADT), and pedestrian counts comprised a full 84.29% of the 12-h count totals,
crash data from the city of Minneapolis are used to build a model of averaged across the 448 included intersections. Of the 448 in the
crash frequencies at the intersection level as a function of modal traffic
inputs. This model determines whether the SIN effect is observable Table 1
within the available datasets for both pedestrians and cars, as well as Hypotheses for signs of variable exponents in per-pedestrian and per-car crash risk
models.
determine specific spatial locations within Minneapolis where pedes-
trians experience elevated levels of risk of automobile crashes, relative Model 6-h peds (bp) AADT (bc)
to intersections elsewhere in the city. The ability to identify specific
unsafe locations based upon aggregated count and crash data offers an Single-variable, pedestrian risk − n/a
Single-variable, car risk n/a −
additional tool for city planners to implement in multimodal and pe-
Multivariable, pedestrian risk − +
destrian-specific planning. Multivariable, car risk + −
2. Methodology
The existence of the SIN effect was examined within collected data Table 2
for the city of Minneapolis, at the intersection level. This framework Dataset summary statistics.
was chosen over other possible areas of analysis, such as mid-block or a
Description Value
link-based framework, due to intersections being the predominant lo-
cation where pedestrians interact with cars. Turning movement counts Intersections with pedestrian counts & AADT 448
(TMCs) for the years 2000–2013 provided pedestrian count data at the Minimum 6-h pedestrians 1
Maximum 6-h pedestrians 14,793
intersection level; AADT measurements from 2000–2013 provided ve-
Average 6-h pedestrians (standard deviation) 832.96 (1843.72)
hicle traffic flow levels on street links; traffic crash records from 6-h peds average percent of 12-h count 82.49%
2000–2013 yielded crash data with location-specific metadata to allow Minimum AADT 252
geocoding; an ESRI shapefile of intersection geolocations in Maximum AADT 40,623
Minneapolis provided by the city allowed for spatial analysis and Average AADT (standard deviation) 8893.33 (5613.00)
Total crashes at sampled intersections 1192 (1180 injuries, 12
geocoding. Datasets pertaining to built environment variables, such as
deaths)
speed limits, number of lanes, and intersection geometry are not in- Minimum crashes at sampled intersections 0
cluded in this analysis. Maximum crashes at sampled intersections 27
A few steps of data processing occurred prior to building models of Average total crashes at sampled intersections 2.66 (3.86)
(standard deviation)
crash counts, and crashes per pedestrian. The TMC data identified in-
dependent pedestrians passing through an intersection, defined by their Note: Summary statistics for datasets used in pedestrian safety analysis: pedestrian
directional heading, across 12-h counts (6 AM to 6 PM); to account for turning movements between 2000 and 2013, AADT between 2000 and 2013, and ag-
this, pedestrian counts for each direction (e.g. northbound, eastbound) gregate crash reports between 2000 and 2013, for the city of Minneapolis.
were summed together to yield a total count. The time windows ex-
tracted from the count data were the three peak periods of the day:
182
B. Murphy et al. Accident Analysis and Prevention 106 (2017) 181–190
sample, 105 intersections had 6-h counts which constituted 100% of the Such a log-linear model form is commonly used within crash mod-
12-h count totals. eling frameworks involving vehicular flows; Abdel-Aty et al. (1998)
The model form used is as follows: give a log-linear framework for modeling crash frequency using de-
mographic and environmental variables, and Lee et al. (2003) give a
b
R = C *Qp p *Qcbc (1) model of crash exposure in terms of vehicular flows and environmental
variables.
where R is the risk factor (either crashes per pedestrian, or crashes per
Single-variable models including only pedestrian traffic or vehicle
car), Qp is 6-h pedestrian traffic flow, Qc is auto traffic flow (AADT), and
traffic, as well as a model with both traffic modes, are included in the
C is a constant. This model form allows for log-linear regression ana-
analysis. Table 1 outlines the hypotheses for the signs of the exponents
lysis, since
in both single-variable and multivariable models of pedestrian and car
log(R) = bp *log(Qp) + bc *log(Qc ) + log(C ) (2) risk factors.
In general, it is hypothesized that increased traffic of a mode has a
183
B. Murphy et al. Accident Analysis and Prevention 106 (2017) 181–190
Histogram of pedestrian counts negative effect on rate of crashes per vehicle or user of that mode (e.g.
increased pedestrians correlate with lower per-pedestrian crash risk);
300
correlates with higher per-pedestrian crash risk). Thus, the SIN effect is
predicted for both pedestrian and auto modes, and it is predicted that
200
modes.
150
100
3. Data analysis
50
Table 2 lists summary statistics for the datasets used in the safety
analysis: automobile-pedestrian crash totals between 2000 and 2013;
0
histograms of the pedestrian counts, AADT data, and crash counts, re-
spectively. Figs. 5 to 7 show maps of the pedestrian counts, AADT data,
Frequency
Lyndale Avenue. All four of these focus areas show a change in pe-
0 5 10 15 20 25
destrian risk, from considering only raw crash counts (Fig. 7) to ac-
Crashes
counting for pedestrian counts (Fig. 10). Areas A and D show elevated
Fig. 4. Histogram of crash count data. pedestrian risk when accounting for pedestrian counts, and areas B and
C show lower levels of per-pedestrian crash risk.
184
B. Murphy et al. Accident Analysis and Prevention 106 (2017) 181–190
4. Regression results levels and the per-automobile risk of hitting a pedestrian. However, the
effect is two orders of magnitude less for automobiles than it is for
Table 3 shows the log-linear regression results for the single-vari- pedestrians. Additionally, an R2 of 0.213 was observed for the single-
able models, and Table 4 shows the log-linear regression results for the variable pedestrian risk model, while an R2 of only 0.098 was observed
multivariate models. For the single-variable log-linear model for SIN for for the single-variable car risk model.
pedestrians, the coefficient of log(6-hpeds) was found to be −0.035 and For the multivariable log-linear model describing per-pedestrian
strongly significant. This is also the exponent bp in Eq. (1), indicating a crash risk, the exponent bp was again found to be negative (−0.036)
negative exponential relationship between pedestrian traffic levels and and strongly significant, indicating a negative relationship between
the per-pedestrian risk of a crash. For the single-variable log-linear increasing pedestrian traffic and per-pedestrian risk of a crash. The
model for SIN for automobiles, the coefficient of log(AADT) was found exponent bc was found to be positive and weakly significant, indicating
to be −0.0003 and strongly significant. This is the exponent bc in Eq. that increased automobile traffic has a positive relationship with the
(1), suggesting a negative exponential relationship between auto traffic per-pedestrian risk of a crash at intersections.
185
B. Murphy et al. Accident Analysis and Prevention 106 (2017) 181–190
Fig. 6. Average annual daily car traffic (AADT). Both dot size
and color scale correlate with AADT levels. (For interpreta-
tion of the references to color in this figure legend, the reader
is referred to the web version of this article.)
Similar relationships were found in the multivariable log-linear per-car models. This is visible in Fig. 8, in which the relationship for
model describing per-car crash risk. Exponent bp was positive and pedestrians appears to be more coherent than that shown in Fig. 9 for
strongly significant, indicating a positive relationship between pedes- cars.
trian traffic and per-car risk of hitting a pedestrian; exponent bc was
negative and strongly significant, indicating a negative relationship 5. Discussion
between auto traffic and per-car risk of hitting a pedestrian. Again, the
model for per-pedestrian crash risk showed a higher R2 (0.219) than The SIN effect was observed in the analysis of the available pedes-
that of the per-car crash risk model (0.117). trian, auto, and crash data at 448 sampled intersections in Minneapolis,
In both single-variable and multivariable models, the SIN effect in using a log-linear model form. The log-linear form was used in lieu of
appeared to be stronger for pedestrians than for cars, indicated by the negative binomial estimators, since the per-user crash rates appeared to
coefficient for log(6-hpeds) in the crashes-per-pedestrian models being follow negative exponential decay, and were also not integer count
larger in magnitude than the coefficient for log(AADT) in the crashes- values. At intersections characterized by higher levels of pedestrian
186
B. Murphy et al. Accident Analysis and Prevention 106 (2017) 181–190
traffic, lower per-pedestrian rates of crashes involving automobiles and known; however, the aforementioned studies have hypothesized psy-
pedestrians were observed; at intersections characterized by higher chological effects on drivers, in that when driving in environments
levels of auto traffic, lower per-car rates of crashes involving auto- characterized by greater average levels of pedestrians, drivers may tend
mobiles and pedestrians were observed. The SIN effect for cars was to act with more caution. And while these various relationships were
found to be a few orders of magnitude weaker than the SIN effect for statistically observed within the traffic data, it is important to note that
pedestrians; SIN pertaining to pedestrian safety is a well-documented the causal directionality of the SIN effect cannot be inferred directly.
phenomenon (see Jacobsen, 2003; Leden, 2002; Bhatia and Wier, The per-pedestrian crash rate was found to increase with increasing
2011). The precise reasons behind this effect are not definitively automobile traffic, and the per-car crash rate was found to increase
187
B. Murphy et al. Accident Analysis and Prevention 106 (2017) 181–190
Crashes per pedestrian vs. average 6−hour pedestrians that there is a ceiling of explanatory power available to a model which
only includes metrics of modal traffic volumes. This limitation is pur-
posefully imposed by focusing the models on only the variables im-
Crashes (2000−2013) per pedestrian
count and crash data is the issue of data quality and availability. Data
practices vary from city to city and state to state, with implications to
investigations intending to aggregate safety data for cross-jurisdiction
5e−03
analysis was 1 over the 6-h count period. The Twin Cities of
Minneapolis and St. Paul are not quite among the most walkable cities
in North America, and rank 16 and 28 in terms of Walk Score, re-
2e−04
188
B. Murphy et al. Accident Analysis and Prevention 106 (2017) 181–190
in each direction. That such roadways would be unsafe for pedestrians its level of pedestrian activity. Pedestrians were found to be at a lower
may seem obvious, but a visualization tool can be more powerful in risk of being hit by a car at intersections with higher pedestrian traffic,
informing planners the scope and extent of dangerous streets within an and individual cars were found to be at a lower risk of hitting pedes-
urban area than simple crash counts alone. trians at intersections with more car traffic. The causality of the SIN
effect is not understood, and more research should be conducted to
6. Conclusion understand its causes, but it is still a justification for improving the
walkability and pedestrian safety of urban environments. Assessing the
Through the pedestrian risk-burden analysis framework outlined in per-pedestrian crash rates at spatial locations, as opposed to crash
this study, it is possible to observe the Safety In Numbers effect at the counts alone, allows practitioners and planners to more readily identify
individual intersection level, for both pedestrians and cars, as well as target areas where improvements to pedestrian infrastructure may be
identify intersections with a disproportionately high rate of crashes for warranted.
189
B. Murphy et al. Accident Analysis and Prevention 106 (2017) 181–190
Table 3 on traffic accident involvement using log-linear models. Accid. Anal. Prev. 30 (6),
Single-variable log-linear regression results. 851–861. http://dx.doi.org/10.1016/S0001-4575(98)00038-4.
Bhatia, R., Wier, M., 2011. “Safety in Numbers” re-examined: can we make valid or
practical inferences from available evidence? Accid. Anal. Prev. 43 (1), 235–240.
log(crashes/ped) log(crashes/car)
Bu, F., Green-Roesel, R., Diogenes, M.C., Ragland, D.R., 2007. Estimating Pedestrian
Accident Exposure: Automated Pedestrian Counting Devices Report. Tech. Rep.
log(6-hpeds) −0.035***
University of California, Berkeley. https://escholarship.org/uc/item/0p27154n.
(0.003) Campbell, B.J., Zegeer, C.V., Huang, H.H., Cynecki, M.J., 2004. A Review of Pedestrian
log(AADT) −0.0003*** Safety Research in The United States and Abroad. Tech. Rep. University of North
(0.00004) Carolina, Chapel Hill.
Constant 0.223*** 0.003*** Do, M.T., Grembek, O., Ragland, D., Chan, C.-Y., 2013. Weighting integration by block
(0.019) (0.0004) heterogeneity to evaluate pedestrian activity. TRB 92nd Annual Meeting. 14 pp.,
schémas, tabl., ill., bibliogr.
Observations 448 448
Greene-Roesel, R., Diogenes, M.C., Ragland, D.R., Lindau, L.A., 2008. Effectiveness of a
R2 0.213 0.098 commercially available automated pedestrian counting device in urban environ-
Adjusted R2 0.211 0.096 ments: comparison with manual counts. TRB 2008 Annual Meeting. pp. 1–16.
Residual std. error (df = 446) 0.109 0.001 Hankey, S., Lindsey, G., Wang, X., Borah, J., Hoff, K., Utecht, B., Xu, Z., 2012. Estimating
F statistic (df = 1; 446) 120.392*** 48.536*** use of non-motorized infrastructure: models of bicycle and pedestrian traffic in
Minneapolis, MN. Landsc. Urban Plan. 107 (3), 307–316. http://dx.doi.org/10.1016/
*** j.landurbplan.2012.06.005.
Note: (standard error); *p < 0.1; **p < 0.05; p < 0.01.
Jacobsen, P.L., 2003. Safety in numbers: more walkers and bicyclists, safer walking and
bicycling. Inj. Prev. 9 (3), 205–209.
Table 4 Jensen, S.U., 2008. Safety effects of blue cycle crossings: a before–after study. Accid.
Multivariable log-linear regression results. Anal. Prev. 40 (2), 742–750. http://dx.doi.org/10.1016/j.aap.2007.09.016.
Leden, L., 2002. Pedestrian risk decrease with pedestrian flow. A case study based on data
log(crashes/ped) log(crashes/car) from signalized intersections in Hamilton, Ontario. Accid. Anal. Prev. 34 (4),
457–464.
log(6-hpeds) −0.036 ***
0.0001*** Lee, C., Hellinga, B., Saccomanno, F., 2003. Real-time crash prediction model for the
(0.003) (0.00002) application to crash prevention in freeway traffic. Transportation Research Board
log(AADT) 0.014* −0.0003*** 82nd Annual Meeting. http://dx.doi.org/10.3141/1840-08.
(0.007) (0.00004) Liu, X., Griswold, J., 2009. Pedestrian volume modeling: a case study of San Francisco.
Constant 0.102 0.003*** Yearb. Assoc. Pacif. Coast Geogr. 71 (1), 164–181. http://dx.doi.org/10.1353/pcg.0.
0030.
(0.067) (0.0004)
Raford, N., Ragland, D., 2004. Space Syntax: An Innovative Pedestrian Volume Modeling
Observations 448 448 Tool for Pedestrian Safety. http://escholarship.org/uc/item/50m064zp.
R2 0.219 0.117 Raford, N., Ragland, D.R., 2005. Pedestrian volume modeling for traffic safety and ex-
Adjusted R2 0.215 0.113 posure analysis: the case of Boston, Massachusetts. Transportation Research Board
Residual std. error (df = 445) 0.109 0.001 85th Annual Meeting.
F statistic (df = 2; 445) 62.284*** 29.462*** Schneider, R.J., Ryznar, R.M., Khattak, A.J., 2004. An accident waiting to happen: a
spatial approach to proactive pedestrian planning. Accid. Anal. Prev. 36 (2),
*
*** 193–211. http://dx.doi.org/10.1016/S0001-4575(02)00149-5.
Note: (standard error). p < 0.1; **p < 0.05; p < 0.01. Schneider, R.J., Henry, T., Mitman, M.F., Stonehill, L., Koehler, J., 2013. Development
and Application of the San Francisco Pedestrian Intersection Volume Model. Safe
All associated datasets and appropriate metadata used within this Transportation Research & Education Center (2).
Score, W., 2016. Walk Score. https://www.walkscore.com/cities-and-neighborhoods/.
study have been uploaded to the Data Repository for the University of Tabeshian, M., Kattan, L., 2014. Modeling nonmotorized travel demand at intersections
Minnesota (DRUM). based on traffic counts and GIS data in Calgary, Canada. Transp. Res. Rec.: J. Transp.
Res. Board. http://dx.doi.org/10.3141/2430-05. 17 pp.
Wier, M., Weintraub, J., Humphreys, E.H., Seto, E., Bhatia, R., 2009. An area-level model
Acknowledgements
of vehicle-pedestrian injury collisions with implications for land use and transpor-
tation planning. Accid. Anal. Prev. 41 (1), 137–145. http://dx.doi.org/10.1016/j.
We thank the University of Minnesota Center for Transportation aap.2008.10.001. http://www.sciencedirect.com/science/article/pii/
S0001457508001930.
Studies, and Roadway Safety Institute, for logistical and financial sup-
Wolshon, B., Wahl, J., 1999. Novi's main street: neotraditional neighborhood planning
port of this research. and design. J. Urban Plan. Dev. 125 (1), 2–16. http://dx.doi.org/10.1061/(ASCE)
0733-9488(1999)125:1(2.
References World Health Organization, 2013. Global status report on road safety. Inj. Prev. 318. ,
http://injuryprevention.bmj.com/content/15/4/286.short. http://www.who.int/
violence_injury_prevention/road_safety_status/2013/en/index.html.
Abdel-Aty, M.A., Chen, C.L., Schott, J.R., 1998. An assessment of the effect of driver age
190