Embed Size (px)
Citation preview
Safety analysis of urban signalized intersections in Kolkata, India 1 using a combined proactive and reactive approach 2
3 4
Submission Date: August 1, 2016 5 Words in Text: 6015 6
Number of Tables/Figures: 6 7 Total word count (including Tables and Figures): 7515 8
9 10 11
Sudeshna Mitra* 12 Associate Professor, Department of Civil Engineering, 13
Indian Institute of Technology Kharagpur, 14 Kharagpur - 721302, India, 15
Phone: +91-9475884524; fax:+9103222-282254 16 Email: [email protected] 17
18 19
Srinivas R. Geedipally, Ph.D., P.E. 20 Texas A&M Transportation Institute 21
Texas A&M University System 22 110 N. Davis Dr., Suite 101 23
Arlington, Texas 76013 24 Phone: (817) 462-0519, fax (817) 461-1239 25
Email: [email protected] 26 27
Dominique Lord 28 Professor, Zachry Department of Civil Engineering 29
Texas A&M University, 3136 TAMU 30 College Station, Texas 77843-3136 31
Phone: (979) 458-3949, fax: (979) 845-6481 32 Email: [email protected] 33
34 (*corresponding author) 35
ABSTRACT 36 Urban signalized intersections in Indian cities report high share of fatalities 37
and serious injuries. However, assessment of safety using rigorous statistical 38 technique is still in early stage in India, primarily due to challenges in obtaining crash 39 data, road geometric and traffic data. This study proposed a methodology combining a 40 proactive and reactive approach by utilizing Road Safety Audit (RSA) principles and 41 statistical assessment techniques, to identify elements posing risk at urban signalized 42 intersections in Kolkata, India. A wide range of factors ranging from the built 43 environment, intersection geometric design and operational parameters are identified 44 and recorded from twenty five sites based on their hypothesised effect from RSA and 45 site visits. Subsequently correlation analysis is performed and Safety Performance 46 Functions are developed using count regression models. 47
Based on the combined proactive and reactive assessments, there is evidence 48 that increased interaction between motorized and non-motorized traffic pose high risk 49 at signalized intersections. Locations, such as markets and commercial areas, 50 busstops, and locations where footpaths are occupied encourage such interactions and 51 witness higher number of crashes. There is also evidence through statistical analysis 52 that signalized intersections with wide major roads are associated with more number 53 of crashes. Finally, results indicate that higher volume of traffic from minor roads 54 compared to major road traffic has diminishing effect on crash risk at the signalized 55 intersections. This unique finding can help plan improved traffic management 56 specifically at those signalized intersections where high volume carrying major roads 57 meets smaller minor roads, for improved safety. 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
INTRODUCTION AND BACKGROUND 86 At-grade intersections are the most complex locations in the transportation 87
network with the maximum numbers of conflicts. Consequently, the identification of 88 elements posing risk of road traffic crashes and the subsequent elimination or 89 relocation of the same are important steps for the safety management of the road 90 network. In doing that, two broad philosophies are often adopted—a) a proactive 91 approach such as Road Safety Audit (RSA) techniques are used to identify potential 92 risky elements and b) a reactive approach where inferences are made by associating 93 the built environment, elements of intersection design and operations with reported 94 crashes. RSA is conducted by an experienced team having knowledge of road design, 95 operation and safety and does not require collection of historical crash data to suggest 96 safety improvements (although they could be collected for specific RSA projects). 97 This approach is widely adopted by practising engineers in many countries around the 98 world. On the other hand, the second category of analysis has been widely used by the 99 safety research community as it involves lesser subjective judgement and utilises 100 scientific methods to identify a set of factors that may influence safety. The Highway 101 Safety Manual (HSM) provides a useful tool in this analysis; even though it requires 102 collection of a sizable data to make a conclusive inference. Safety Performance 103 Function (SPF) is one such method, which has been extensively discussed in HSM, 104 that has been used by researchers for safety analysis for over two decades. 105
SPFs are mathematical relationships that associate crash frequencies with 106 geometric design elements, traffic volume and regulatory factors; and are used either 107 for better understanding crash risk or for prediction purpose. A significant proportion 108 of past studies (1 - 8) have focused on modeling accidents at signalized intersections, 109 primarily due to high occurrence of crashes at urban signalized intersections. Ideally, 110 if the SPFs are specified correctly, they should help provide insights into safety by 111 identifying factors explaining highest variability in crash occurrence at urban 112 intersections. Thus, in developing SPFs, researchers first identify a set independent 113 variables having high correlation with the response variable, i.e. frequency of crashes 114 in this case. However, correlation coefficients show linear relation between two 115 variables, and it is possible that two variables may have non-linear relationship 116 between them. 117
As opposed to developed countries, there have been fewer studies for 118 developing countries where road environment and built environment surrounding 119 roads are very different. This is probably due to the fact that obtaining relevant data 120 for safety assessment and the development of SPF is more challenging in developing 121 nations. This is because the data collection and maintenance are often not done 122 systematically. In India, crash data are still collected and recorded manually using 123 First Information Recording (FIR) form except in three southern states. This requires 124 extensive post processing. Furthermore, police reported data consist of significant 125 underreporting of less injury crashes, such as property damage only and minor injury. 126 Obtaining traffic volume data also requires significant resources, which often limits 127 the size of sample. A small sample of network locations also poses bias in model 128 estimation as was shown by Lord (9). Due to such limitation in data availability in 129 general and crash data in particular, few research organizations in India have made 130 efforts to conduct safety assessment of transport network utilising historical crash data 131 (10). However, because of these issues with data, RSA based safety assessment has 132 gained popularity and momentum and faced lesser challenges in execution. While 133 RSA during the design and construction phases probably has no competing 134 alternative, RSA during operation or RSA findings from existing roads should be 135
cross checked with historical evidence wherever possible. This demands a need to 136 explore a possible alternative way by combining both reactive and proactive 137 approaches using which reasonable assessment of safety, identification of hazardous 138 elements and future prediction of safety are possible. 139
To that end, the primary objectives of this study are to determine 140 a) if there exists any significant association between elements of built environment 141 such as land use, infrastructure, geometric design, and traffic operation parameters 142 which are obtained from RSA and site visit with occurrence of total crashes at urban 143 signalized intersections, 144 b) quantification of the effects of those factors which influence crash occurrence at 145 urban signalized intersections by developing a SPF. 146
To achieve these objectives, this study utilises concepts of RSA techniques to 147 identify elements of built environment and design that may pose safety hazard, 148 perform a detailed correlation analysis to identify association of various factors with 149 total crashes at signalized intersections, and finally develop an appropriate crash 150 prediction model to quantify the effects of these factors on safety of signalized 151 intersections. The remainder of the paper is organised as follows: a literature review is 152 given first, followed by the methodology, description of data, results and conclusion. 153 154 LITERATURE REVIEW 155
There have been extensive studies in developed and some in developing 156 countries on safety assessment of signalized intersections. The literature in this regard 157 can be divided into two aspects, a) in terms of modelling methodology used and b) in 158 terms of identification of factors influencing intersection crashes. 159
Count data models such as Poisson and NB regression models are commonly 160 used in crash data modeling and analysis. Although count data modeling always starts 161 with Poisson regression, Poisson formulation is often not suitable in crash modeling 162 due to the over dispersed (i.e. variance > mean) nature of crash data. Negative 163 Binomial models are generally used for such cases as it includes a gamma-distributed 164 error term in Poisson mean to account for a wide range of unobserved heterogeneity 165 such as omission of relevant variables, measurement error, or just the intrinsic 166 randomness in count data. Both Poisson models (11-13) and negative binomial (NB) 167 models (1, 2, 14-17) have been used to investigate relationships between crash 168 occurrence and intersection characteristics. In developing models, one of the major 169 concerns is estimation bias that may arise from model misspecification and 170 unobserved heterogeneity. As shown by Mitra and Washington (18), in the presence 171 of a well-defined mean function, the extra-variance structure of NB generally 172 becomes insignificant, which helps in better inference by reducing standard error of 173 estimation and forming a narrow confidence interval. Other possible sources of bias 174 are omitted variable bias (Mitra and Washington, 2012, (19) bias due to small sample 175 (9) and bias due to unobserved heterogeneity (20-22) which have been addressed by 176 the safety research community. However, traditional negative binomial regression 177 models are used as a starting point to model overdispersed crash data and it’s 178 variation such as random parameter negative binomial models have been widely used 179 by the researchers to capture unobserved heterogeneity present in the data. 180
Regarding identification of factors affecting the safety of signalized 181 intersections, previous research on intersection related risk factors (e.g., 1-7) has 182 shown that certain variables including traffic volume (e.g., total volume and right-turn 183 volume), geometric design (e.g., intersection type and number of lanes) and traffic 184 control (e.g., phase number and signal control type) can all have significant impacts 185
on signalized intersection safety. Factors such as lane width (4), presence of median 186 or median width (2,14), presence of turning lane (8), type of signal and number of 187 phases (3), land-use of the site (14), presence of bus stop (2), volume of entering 188 traffic, i.e. major and minor road volume or some variable to capture the effect of 189 these variables (5-8), have been found to be associated with crash occurrence at 190 signalized intersections either in positive or negative manner. There has also been 191 research on built environment and its effect of road safety (4, 23, 24). Built 192 environment refers to the manmade surroundings that provide the setting for human 193 activity, ranging in scale from buildings, offices and parks or green space to 194 neighbourhoods and cities. Thus built environment is represented by land-use 195 characteristics, transport infrastructure and road network characteristics as well as 196 demographics. The literature suggests that the built environment affects crash 197 frequency and severity through the mediators of traffic volume and traffic speed (23). 198 Findings from developed countries suggest that development patterns in urban built 199 environment impact safety primarily through the traffic volumes they generate, and 200 secondarily through the speeds they encourage, which in turn affect frequency and 201 severity of crashes. Further, not only traffic speed but the speed differentials in the 202 traffic stream and the conflicting traffic movement affect crash occurrence. The 203 conflicts are again greatly dependent on land-use, access management, transit and 204 pedestrian infrastructure. In the context of developing countries, role of design built 205 environment is expected to be significant as certain design encourage greater access, 206 resulting in more interaction between motorized and non-motorized traffic. Higher 207 interaction between motorized and non-motorized traffic also affects safety of 208 motorized vehicles, due to their sudden evasive manoeuvre to avoid collisions with 209 non-motorized traffic. As a result, findings from US, UK, or even China may not be 210 extendable for Indian condition due to the very different traffic conditions 211 characterised by higher interaction between slow and high speed vehicles as well as 212 pedestrians. 213 214 METHODOLOGY 215
A four step methodology was followed combining proactive and reactive 216 approach to identify factors affecting the safety of signalized intersections in Kolkata, 217 India. The steps are discussed in this section and are illustrated in Figure 1. It includes 218 site visit to conduct an RSA, identification and collection of elements of road design 219 and traffic operations, performing a correlation analysis and finally development of a 220 regression model to identify factors affecting safety of signalized intersections in 221 Kolkata. 222 223 224 225 226 227 228 229 230 231 232 233 234 235
236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 Figure 1 Flow chart showing methodological steps followed in safety assessment 263 of signalized intersections in Kolkata. 264 265 Road safety audit 266
A RSA is a formal procedure for assessing crash potential and safety 267 deficiency of an existing or future road by an independent audit. This is viewed as a 268 proactive low-cost measure to improve safety. RSA is a systematic way of checking 269 the road alignment, various cross sectional elements, road side design and 270 appurtenances as well as traffic operational and management measures that will 271 ensure safety of all road users. Hence, a major aim of RSA of a road under operation 272 is to identify elements that may increase risk of road traffic crashes and elimination 273 and/or relocation of the same. In conducting an RSA of existing road network, certain 274 checklists are followed. Since the work involved in this study focuses on safety of 275 signalized intersections, the list shown in Table 1 was used. Another issue with safety 276 reviews of existing roads revolves around the use of crash data. Some auditors like to 277 have access to the Police crash data for the road they are reviewing. They say it helps 278 them understand some of the proven safety issues along the road. Others argue that 279 this can cause the review team to focus too closely on the crash sites, possibly 280 overlooking other high risk locations. Nonetheless, use of crash data, if available, 281 definitely provides some insight about the underlying safety of the location which 282 otherwise would be missed if they are not available. 283 284 285
Correlation analysis
Identification of Signalized intersections based on crash data collected from Kolkata Police
Road Safety Audit of Signalized Intersection
Development of SPF
Overdispersion tests
Identification of suitable count data models
Final Identification of factors from SPF
Identification of hazardous elements for signalized intersection
Data collection
286 Table 1 Checklist for Road safety Audit for Signalized intersection 287 288
General Signal-Controlled Junction 1) Is the general layout of junction caters safely for all road users including disabled road users? 2) The type of junction (T-type, staggered, signal controlled, roundabout) suitable for the function of the two or more roads, the traffic volume, the traffic movements (pedestrians and vehicular) and the site constraints? Is it safest alternative? 3) Is the layout of the junction adequate for all permitted vehicular movements and for all types of vehicles? 4) Is there adequate provision for channelizing the different streams of traffic? (Check the provision for right turn lanes, deceleration lanes and acceleration lanes?) 5) Is adequate provision made for pedestrians and non-motorised vehicles? 6) Is the provision of night-time lighting adequate, if not what are the deficiencies? 7) Are junction(s) at that stretch having proper markings, signs and studs to avoid accidents?
1) Does the signal sequence conform to the requirements of the regulations and standards? 2) Do the signals clearly indicate which movements are allowed at any one time? Are the timings of various phases of signal cycle adequate? 3) Are the signals for competing phases located in such a way that they are visible only to the traffic for which they are intended? 4) Are all right turning movements protected as far as possible? 5) Does the signing; marking and channelization make it clear to drivers what path they should take through the junction? 6) Are pedestrian crossings places marked, and are pedestrians channelled to these crossings? 7) Whether the pedestrian crossing signal controls are provided where appropriate? If so, there is a need for the crossing movements to be fully protected from conflicting traffic movements for example where there will be serious conflicts with turning traffic.
289 Data collection 290
Crash data for the 25 signalized intersections were obtained from Kolkata 291 Traffic Police. Kolkata police has divided the entire Kolkata urban area in twenty five 292 traffic guards for better management. Out of these traffic guards, this study focused 293 on selected intersections in nine traffic guards where at least one injury crash occurred 294 between 2011 and 2014. The traffic guards were Beliaghata, Howrah Bridge, 295 Shyambazar, Ultadanga, Vidyasagar Setu, Sealdha, Head Quarter, Joradagan, and 296 South. All 25 signalized intersections had a video camera equipped, from which 297 traffic volume data was extracted. 298
Regarding land-use and design data, a separate site visit was performed after 299 the preliminary audit of the sites. Collected data include information about road type 300 (arterial or minor road), land-use pattern in percentages (residential, commercial, 301 school zone, government office, park/recreational, open space), pavement condition 302 (good, raveling, depressed side drains, pot hole), geometric features, such as straight 303 or curved road, carriageway width, sight distance (clear/obstructed), percentage of 304 carriageway blocked and width of median. Type of traffic operation, such as one way 305 or two way, junction type such as four legged or three legged were collected. Built 306 environment variables, such as presence of pedestrian facilities, zebra crossing, foot 307 over bridge, subway, presence of metro station, bus stop, pedestrian facility at bus 308 stop, central refuse island, zebra crossing at median opening, zebra crossing width 309 foot path width, percentage of foot path occupied, stop marks, road marking visibility, 310 presence of tram, presence of street light were also assembled. The summary statistics 311 of these variables are given in Table 2. 312 313 314 315 316
317 Table 2 Descriptive Statistics 318
Min Max Mean
Std. Deviation
Total_crashes 1 16 4.56 3.990
Major_road_carraigeway_width (m) 7 11 8.96 1.773
Minor_road_carraigeway_width (m) 7 11 7.30 .968
Major_road_volume (Veh) 4855.0 38086.0 19380.2 7639.8
Minor_road_volume (Veh) 3393.0 21380.0 9880.0 4394.9
Ratio_major_minor_volume 0.2 1.0 0.6 0.2
Ped_volume_peak (number) 546.0 6974.0 2666.0 1461.6
Busstop (0/1) 0 1 0.20 0.408
Metrostation (0/1) 0 1 0.12 0.332
Tram_line_location (0/1) 0 1 0.32 0.476 Presence_road_marking 0 1 0.28 0.46 Presence_stop_marking 0 1 0.64 0.49
Footpath_width (m) 1.86 3.88 2.6936 0.45812
Footpath_occupancy_proprtion (%) .0 .5 0.160 0.1683
Land_use_residential (%) .0 .5 0.244 0.1734
Land_use_commercial (%) .1 .8 0.476 0.1964
Land_use_goverment_office (%) 0 1 0.17 0.214
Land_use_open_space (%) 0 1 0.06 0.138
319 Correlation analysis 320
To start with the modelling process, it is important to know the correlation 321 structure present in the data. For this purpose, linear correlation using Pearson’s 322 correlation parameter test is often used. However, if the data are measured on the 323 ordinal or categorical scale, or any data on continuous scale may not satisfy the 324 requirement of approximate normality (25), Spearman rank correlation parameter 325 should be used. Independent variables that are highly correlated with dependent 326 variable (i.e. crash frequency in this case) should be included in the model and two 327 independent variables with high correlation should not be included together in the 328 model. However, two variables with low correlation coefficient may still be correlated 329 in a nonlinear manner. 330
Once the variables are identified, first of all variables that have high 331 correlation with dependent variable, i.e total crash frequency are included separately 332 to assess their explanatory power. After that, an incremental model building was 333 adopted by adding and dropping variables to arrive at the final model. 334
335 Development of statistical regression model 336
Over the last two decades, there has been significant research into and 337 improvement of count data modelling methodologies for crash frequency data (18, 21, 338 26, 27, 28). 339
In a Poisson regression model, the probability of intersection i having in340
accidents (where in is a non-negative integer) is given by: 341
!
)exp()(i
ini
i nnP
i µµ −= (1) 342
Where )( inP is the probability of intersection i having in accidents and iµ is the 343 Poisson parameter for intersection i, which is equal to intersection i's expected 344 number of accidents, ][ inE . In general, Poisson regression models are estimated by 345
specifying the Poisson parameter iµ (the expected number of events per period such 346
as per year) as a function of explanatory variables kXXX .........., 21 such that 347
(2) 348
Here β is the vector of estimable coefficients. In this formulation, the expected 349 number of events per period is given by 350 )exp(][ iii XnE βµ == (3) 351
Further, Poisson model assume mean and variance equality, i.e. E[ in ] = VAR[352
in ]. However, for crash data, this assumption often does not hold, i.e the data either 353
under-dispersed (E[ in ] >VAR[ in ]) or over-dispersed (E[ in ] <VAR[ in ]), and the 354 coefficient vector will be biased if inequality of variance is not taken into account 355 (Mitra and Washington, 2007). 356
Cameron and Trivedi (29) have proposed a number of tests to check for under 357 or overdispersion in the Poisson regression model. The testing framework is built 358 around the following hypotheses (30). 359 𝑁𝑢𝑙𝑙 ℎ𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠: 𝐻0:𝑉𝑎𝑟 [𝑦𝑖] = 𝜇𝑖 (4) 𝐴𝑙𝑡𝑒𝑟𝑛𝑎𝑡𝑖𝑣𝑒 ℎ𝑦𝑝𝑜𝑡ℎ𝑒𝑠𝑖𝑠: 𝐻1:𝑉𝑎𝑟 [𝑦𝑖] = 𝜇𝑖 + 𝛼𝑔(𝜇𝑖) (5) 360
One of the main assumptions behind this test is that the Poisson model gives 361 consistent estimates of E [𝑦𝑖] = 𝜇𝑖 and result in a model where E [𝑦𝑖] = Var[𝑦𝑖] . 362 However, in case of overdispersed count data Negative Binomial (NB) model is used, 363 which is derived by adding an additional term in variance as shown in equation 5. 364 Here, 𝛼 is the overdispersion parameter and follow a gamma distribution. The test is 365 carried out by testing the significance of the single coefficient in the linear OLS 366 regression of 367 𝑧𝑖 = [(𝑦𝑖 − 𝜇𝑖)2 − 𝑦𝑖]/(𝜇𝑖√2) (6) 368 on 369 𝑤𝑖 = 𝑔(𝜇𝑖)/𝜇𝑖√2 (7) 370
Two possibilities are suggested: 371 𝑔(𝜇𝑖) = 𝜇𝑖 𝑎𝑛𝑑 𝑔(𝜇𝑖) = 𝜇𝑖2 (8)
A significant 𝛼,𝑔(𝜇𝑖) indicates that if 𝛼 𝑎𝑛𝑑 𝑔(𝜇𝑖)are significant, and that H0 372 should be rejected in favor of H1. This implies that the count is overdispersed and that 373 negative binominal regression should be adopted. 374
Finally, to measure the overall goodness of fit, the log-likelihood ratio index is 375 calculated. For count data models, the common practice is to use a “pseudo” R2 376 statistic, which is often known as log-likelihood ratio index 2ρ . According to 377
Washington et al. (2010), 2ρ is given by 378
)0()(12
LL β
−=ρ (9) 379
380 RESULTS AND DISCUSSION 381
Based on the RSA, the major element of risk identified was high interaction of 382 motorized and non-motorized traffic, such as pedestrians, especially near market 383 areas, commercial areas and bus stops. Road conditions were also found to be poor in 384 these locations as opposed to locations where government offices are present. Road 385
ikkiii XXX ββββµ ++++= ...............)log( 22110
marking was another major problem primarily in residential areas. Signalized 386 intersections in residential areas were also found to have narrower roadway and 387 sidewalk width. A common observation across all locations is poor usage of marked 388 pedestrian crossing. Parking by the side of road was observed only at commercial 389 areas. Major observations from RSA are shown in Figure 2. 390
Zebra Crossing is occupied by the car, Pedestrians are not
following the proper path
Pedestrians are not using footpath
Illegal crossing of pedestrian
Pedestrians are not using Zebra Crossing
Pedestrian behaviour at high commercial zone
Zebra Crossing is occupied by buses, Pedestrians are not
following the proper path
Bus stop at intersection
Pedestrians at bus stop
Figure 2 Findings from Road Safety Audit across various study intersections. 391
392
The statistical assessment in this study started with a correlation analysis using 393 Spearman Rank Correlation test to estimate correlation among the independent 394 variables and that with the dependent variable. The results are shown in Table 3. As 395 can be seen, the total crashes at signalized intersection was highly correlated with 396 presence of bus stop (0.55, p=<0.01) and width of major road (0.418, p<0.05). On the 397 other hand, the correlation with minor road width (-0.09, p>0.2) was weak with total 398 crashes. Traffic volume from major road, minor road, and variables capturing the 399 effect of built environment, such as residential and commercial areas had some 400 correlation with the dependent variables, but none of them were significant at least at 401 5% level of significance. Further, none of the independent variables had statistically 402 significant high correlation among them, expect the width of minor road and 403 residential land-use (0.5, p<0.01). But these two variables were not found to be 404 significant in the final model. A variable capturing the ratio of minor road volume to 405 major road volume was also created and this variable definitely has high correlation 406 with major road and minor road volume, which dictates that the major road volume 407 and the ratio of major to minor road volume should not be included at the same time 408 in model building. Variables capturing the effect of built environment, such as 409 footpath width and their occupancy showed very interesting results. Footpath widths 410 were found to be statistically significantly wider in areas where government offices 411 are located (0.62, p<0.01), but narrow in residential areas (-0.58, p<0.01). Footpath 412 width is also negatively correlated with commercial areas, but the correlation is not 413 statistically significant. On the other hand, it was interesting to identify that footpath 414 occupancy (occupied illegally by street vendors) is statistically significantly higher in 415 commercial areas (0.6, p<0.01) and significantly lower in areas where parks (-0.5, 416 p<0.01) are located. Narrow sidewalk width and their high encroachment are factors 417 posing risk to non-motorized road users in commercial areas. Presence of road 418 marking was found to be high in areas where government offices are located 419 compared to residential and commercial areas, but none of these correlations were 420 statistically significant. Finally, the variable capturing traffic volume- the ratio of 421 minor road traffic to major road traffic, found to have high correlation with minor 422 road carriageway width (0.41, p<0.05). As a result, these two variables are not 423 included together in regression model. 424
Table 3 Results from the correlation analysis 425
Tota
l_cr
ashe
s M
ajor
_roa
d_ca
rraig
eway
_wid
th
Min
or_r
oad_
carra
igew
ay_w
idth
Maj
or_r
oad_
volu
me
Min
or_r
oad_
volu
me
Rat
io_m
ajor
_min
or_v
olum
e
Ped_
volu
me_
peak
Bus
stop
Met
rosta
tion
Tram
_lin
e_lo
catio
n
Pres
ence
_roa
d_m
arki
ng
Pres
ence
_sto
p_m
arki
ng
Foot
path
_wid
th
Foot
path
_occ
upan
cy_p
ropr
tion
Land
_use
_res
iden
tial
Land
_use
_com
mer
cial
Land
_use
_gov
erm
ent_
offic
e
Land
_use
_par
k
Land
_use
_ope
n_sp
ace
Total_crashes 1 Major_road_carraigeway_width .418* 1 Minor_road_carraigeway_width -.09 .327 1 Major_road_volume .203 .123 -.15 1 Minor_road_volume -.09 .067 .32 .08 1 Ratio_major_minor_volume -.11 .089 .41* -.5** .6** 1 Ped_volume_peak .137 .263 -.08 .31 .01 -.32 1 Busstop .55** .242 .14 .33 .26 .07 .03 1 Metrostation .104 .327 .60** -.08 .38 .29 .31 .12 1 Tram_line_location .072 -.25 .30 -.26 .30 .29 -.12 .30 .27 1 Presence_road_marking .106 .014 .05 -.18 -.22 .10 -.12 .13 .04 -.05 1 Presence_stop_marking .257 .17 .04 .12 -.06 .01 -.02 .38 .02 .16 .46* 1 Footpath_width -.24 .10 .49* -.29 .13 .33 -.21 -.09 .25 .08 .26 .02 1 Footpath_occupancy_proprtion -.05 -.14 .09 .14 .21 .05 .33 -.12 .42* .50* -.15 .2 -.3 1 Land_use_residential .186 .17 -.5** .15 -.26 -.36 .19 .00 -.30 -.18 -.09 .0 -.58** .13 1 Land_use_commercial .177 -.09 -.06 .12 .12 -.06 .17 -.01 .27 .37 -.21 -.06 -.23 .6** .04 1 Land_use_goverment_office -.09 .07 .55** -.12 .10 .31 -.05 .19 .21 .07 .36 .10 .620** -.26 -.6** -.42* 1 Land_use_park -.292 -.16 -.18 -.29 -.24 -.10 -.20 -.25 -.18 -.34 .19 -.20 .26 -.5** -.18 -.51** .17 1
Land_use_open_space .005 -.07 -.20 -.06 -.21 -.01 -.21 -.28 -.21 -.38 .00 -.19 -.04 -.35 -.19 -.34 -.03 .14 1 *significant AT 95% C.I **significant at 99% C.I
The next step was to test for over-dispersion in crash data for which the test 426 suggested by Cameron and Trivedi (29) was conducted and the result indicated 427 presence of over-dispersion in crash data (Refer Table 4). After this diagnosis, 428 individual variables are added separately to check their statistical significance in 429 explaining the dependent variable. In doing that, the overall model prediction 430 capability by these variables as well as the variable’s ability to correctly specify the 431 model was tested. The results of this exercise are given in Table 4. As can be seen in 432 Table 4, the model explanatory power by adding the variable “Busstop” was good, but 433 the dispersion parameter resulted in an insignificant estimate. Further, the test for 434 over-dispersion indicated absence of over-dispersion in the data, which violates the 435 fact. This is probably due to large variability in the variable bus-stop as it is present 436 only in 20% of locations in the dataset. Further, due to small sample size, it was not 437 possible to drop the locations with bus stop and rerun a model to check the effect of 438 other variables. While bus stop was decided to be excluded for model prediction 439 purpose, the presence of bus stops definitely increases the number of conflicts among 440 the mixed traffic when vehicular speeds are more heterogeneous, as was pointed by 441 Wong et al. (4). Field observation of the same shows very similar evidence. In 442 Kolkata city, most bus stops are classified as nearside just ahead of signalized 443 intersections as can be seen in Figure 2. In most cases, passengers wait on the road 444 and there is no bus bay present at the site. As a result, there is higher interaction with 445 passengers waiting on the road and with other motorized vehicles, which definitely 446 pose high safety hazard. Hence, even though this variable was not included in the 447 final model due to data limitation; there is evidence that this factor related to design 448 built environment of Kolkata city poses hazard at the signalized intersection. This 449 conclusion was achieved by a combined correlation analysis (Refer Table 3) and by 450 safety audit of the bus stops located at signalized intersections in Kolkata city. 451
The other variables, such as width of major road, residential and commercial 452 land-use, ratio of minor road and major road traffic volume were all found to be 453 significant when added independently and they all resulted in correct model 454 specification. Once this preliminary investigation was performed, these independent 455 variables are incrementally added or deleted and the final model resulted in a negative 456 binomial model denoted as Model 7 and the results are shown in Table 4. Here it is 457 important to mention that advanced model such as random parameter negative 458 binomial model was not used due to small sample size. Results show that Commercial 459 area (β = 1.19, p < 0.100) and wider width of major road (β = 0.228, p< 0.006) 460 increase total crashes whereas a higher ratio of volume of minor road traffic compared 461 to major road traffic volume (β = -0.766, p< 0.118) decreases total crashes at 462 signalized intersections. 463
Table 4 Results from count data models 464 465 466
467 468
Attributes NB regression Model-1
(Constants only)
NB regression Model-2
(bus-stops)
NB regression Model-3
(Residential)
NB regression Model- 4
(commercial)
NB regression Model-5 (Width)
NB regression Model-6 (Volume
Ratio)
NB regression Model-7
(Combined)
Model Coefficients (t-stat) Constant 1.51 (9.45) 1.19 (7.66) 1.1(3.86) 1.06 (2.51) -0.45 (-0.55) 1.94 (4.92) -0.79 (-0.8) Busstops 1.07 (3.74)
Residential 1.56 (1.70) Commercial 0.91 (1.13) 1.19 (1.64)
Major Carriageway Width 0.21 (2.42) 0.23(2.73) Ratio of minor road traffic to
major road traffic -0.777 (-1.19) -0.766 (-1.35)
Results from overdispersion Test g=mu(i): 2.19 1.49 2.63 2.03 2.48 2.12
g=mu(i)^2: 2.19 1.33 3.11 2.03 2.88 2.21 Dispersion parameter for count data model
Alpha (α) 0.426 (2.41) 0.18 (1.55) 0.36 (2.25) 0.39 (2.23) 0.31 (2.11) 0.39 (2.33) 0.23 (1.85) Overall Goodness-of-fit
Log-Likelihood function -63.15 -57.59 -61.71 -62.52 -60.47 -62.46 -58.39 Restricted Log-Likelihood
function -74.89 -74.89 -74.89 -74.89 -74.89 -74.89 -74.89
ρ2 0.157 0.231 0.175 0.165 0.193 0.166 0.22
The variable percentage of commercial area near the signalized intersection is 469 another variable capturing the effect of design built environment where a signalized 470 intersection is located. Generally, commercial areas are characterized by increased 471 access which increases frequent interaction of vehicles. These locations also 472 experience high road side commercial activities. Further, the results from the 473 correlation analysis indicated that commercial areas in the dataset are also associated 474 with high footpath occupancy, which probably forces pedestrians to walk and engage 475 in commercial activities along the road, resulting in frequent interaction between 476 pedestrian and motorized vehicles, leading to higher risk of crashes. 477
The variable carriageway width of major road or simply width of major road 478 found to be highly significant and positively associated with total crashes. This is not 479 surprising as higher road width encourages higher speed and greater freedom in 480 vehicle maneuver which leads to poor safety. Signalized intersections in Kolkata are 481 found to be no exception and this finding empirically proves that wider roads in urban 482 area pose safety hazard. However, width of minor road was not statistically 483 significant and in the correlation analysis, a wider minor road width was actually 484 found to be negatively associated with total crashes, even though the variable was not 485 statistically significant. Effect of width of road however, found to be mixed in earlier 486 research. 487
The most interesting finding in this study is the effect of traffic volume. From 488 the correlation analysis, while the traffic volume from major road was somewhat 489 correlated with total crashes in a positive manner, traffic volume from minor roads 490 had very low and negative correlation with total crashes. After incorporating the 491 logarithmic transformation of traffic volume, only major road traffic volume showed 492 some significance and minor road traffic volume was not at all significant. However, 493 in road traffic safety literature, major and minor road traffic volumes are always 494 considered to be the most important variables in intersection crash prediction as they 495 capture exposure. With this in mind, a different form of total entering volume, the 496 ratio of minor road traffic to major road traffic was included in the model and resulted 497 in negative correlation with total crashes. While it may appear to be strange, the width 498 of minor road and the minor road traffic volume were consistently found to be 499 negatively associated with total crashes even though they were not significant. Thus, 500 the effect of minor road traffic compared to the major road traffic was tested and the 501 result indicates that signalized intersections in Kolkata with higher minor road traffic 502 compared to major road volume are safer than intersections where less volume 503 carrying minor road meets with a major road carrying higher volume of traffic. This 504 probably is a surrogate capturing the priority of a signalized intersection from the 505 traffic operations and management point of view, which is quite localized and unique 506 for many Indian cities like Kolkata. Safety Audit and site visit shows that major 507 signalized intersections (often intersections of two roads of equal widths) are better 508 managed with the help of traffic police who enforce and ensure orderly movement of 509 traffic than smaller intersections. As a result, this finding is not against expectation 510 but is a revelation made possible by statistical modeling. While the other two findings 511 indicate safety effects of built environment and design, this finding can directly help 512 Kolkata police to further review safety of such signalized intersections and plan for 513 better traffic management. 514
Finally, it is important to point out that even though only three variables were 515 found to be statistically significant in the final model, the casual contribution to 516 crashes at intersections may not be fully inferred from a regression analysis, as 517 pointed by Wong et al. (4). This is more so when the sample size is small. However, 518
due to practical difficulty of data collection and obtaining crash data in a country 519 where systematic crash data collection is absent, this study is an attempt to study 520 associations between signalized intersection related variables and total crashes. 521 Identification of risky elements associated with urban signalized intersection through 522 the correlation analysis, safety audit by site visit and finally findings from regression 523 model provide some reasonable directions for improvement and further exploration. 524 525 CONCLUSION 526 Road traffic safety in urban India is a matter of serious concern. This is particularly 527 true in major cities such as Kolkata. However, due to the absence of systematic crash 528 data collection and maintenance, there have been limited studies on traffic safety of 529 urban India. With the exception of only one study (31), existing studies (32) on road 530 safety in the context of urban India either follow road safety audit techniques or 531 mostly a descriptive approach where overall trends of safety are documented. 532 However, no studies to-date focused on spatial location of the intersection by 533 capturing the elements of built environment, geometric design and traffic operations 534 on safety in the Indian context. In that sense, this study makes a unique attempt to 535 collect traffic exposure and a wide range of design built and facility design related 536 factors to investigate their effects on safety of signalized intersections in Kolkata, 537 India. 538
While safety assessment of road network has been widely conducted using 539 crash data analysis and development of crash prediction models, this study has 540 proposed a methodology combining a proactive and reactive approach by utilising 541 Road Safety Audit principles and statistical assessment techniques involving 542 correlation and regression analysis. For this purpose, crash data from Kolkata city was 543 collected and a total of twenty five signalized intersections were visited for safety 544 audit. Videographic data were also obtained for these intersections from Kolkata 545 Police. 546
Based on the combined proactive and reactive analyses it was concluded that 547 increased interaction between motorized and non-motorized traffic is definitely a 548 major cause of risk at signalized intersections. Locations, such as markets and 549 commercial areas, busstops, and locations where footpaths are occupied encourage 550 such interactions and evidence from either correlation analysis or statistical model 551 also show similar results. As was mentioned earlier, presence of bus stop was highly 552 significant both from correlation analysis and preliminary count model, but this 553 variable had to be dropped due to its small sample size affecting model specification. 554 However, commercial built-up areas was found to be significant and included in final 555 model for prediction purpose. The other important finding through statistical analysis 556 was width of major road, which positively affected total crashes. Finally, ratio of 557 traffic from minor and major road provides an interesting insight into safety of the 558 selected intersections in this study. 559
Like any other study, this study is not without limitation. A major concern was 560 small sample size, which could affect estimation. Further, model validation and 561 transferability for other intersections in the Kolkata have not been tested yet. Thus, an 562 extended study is required to address these concerns by reinvestigating crash 563 prediction model for signalized intersections in Kolkata by incorporating more sites 564 and road users’ behavior. Nonetheless, this study provides important insights into the 565 safety of signalized intersections in a major urban area in India. 566
567 ACKNOWLEDGEMENT 568
The authors would like to acknowledge the help of Mr. T. Subudhi, Mr. Sanil Jain and 569 Mr. Satyajit Rana for data collection and processing. 570 571 572 573 REFERENCES 574
1. Poch, M., and F. Mannering. Negative binomial analysis of intersection-accident frequencies. 575 Journal of Transportation Engineering, Vol. 122, No. 2, 1996, pp. 105-113. 576
2. Chin, H. C., and M. A. Quddus. Applying the random effect negative binomial model to 577 examine traffic accident occurrence at signalized intersections. Accident Analysis & 578 Prevention, Vol. 35, No. 2, 2003, pp. 253-259. 579
3. Kumara, S.S.P., and H. C. Chin. Modeling accident occurrence at signalized tee intersections 580 with special emphasis on excess zeros. Traffic Injury Prev. Vol. 3, No. 4, 2003, pp. 53–57. 581
4. Wong, S. C., N. N. Sze., and Y. C. Li. Contributory factors to traffic crashes at signalized 582 intersections in Hong Kong. Accident Analysis & Prevention, Vol. 39, No. 6, 2007, pp. 1107-583 1113. 584
5. Wang, X., M. Abdel-Aty., A. Nevarez., and J. Santos. Investigation of safety influence area 585 for four-legged signalized intersections: nationwide survey and empirical inquiry. 586 Transportation Research Record: Journal of the Transportation Research Board, (2083), 2008, 587 pp. 86-95. 588
6. Guo, F., X. Wang., and M. A. Abdel-Aty. Modeling signalized intersection safety with 589 corridor-level spatial correlations. Accident Analysis & Prevention, Vol. 42 No. 1, 2010, pp. 590 84-92. 591
7. Gomes, S. V., S. R. Geedipally., and D. Lord. Estimating the safety performance of urban 592 intersections in Lisbon, Portugal. Safety science, Vol. 50 No. 9, 2012, pp. 1732-1739. 593
8. Xie, K., X. Wang., H. Huang., and X. Chen. Corridor-level signalized intersection safety 594 analysis in Shanghai, China using Bayesian hierarchical models. Accident Analysis & 595 Prevention, Vol. 50, 2013, pp. 25-33. 596
9. Lord, D. Modeling motor vehicle crashes using Poisson-gamma models: Examining the 597 effects of low sample mean values and small sample size on the estimation of the fixed 598 dispersion parameter. Accident Analysis & Prevention, Vol. 38, No. 4, 2006, pp. 751-766. 599
10. Raban, M, Z., L. Dandona., R. Dandona. The quality of police data on RTC fatalities in India. 600 InjPrev, Vol. 20 No. 5, 2014 pp. 293–301. 601
11. Jones, B., L. Janseen., and F. Mannering. Analysis of the frequency and duration of the 602 freeway accidents in Seattle. Accident Anal. Prevention Vol. 23, No. 4, 1991, pp. 239–255. 603
12. Joshua, S., and N. Garber. Estimating truck accident rate and involvement using linear and 604 Poisson regression models. Transportation Planning and Technology, Vol. 15, No. 1, 1990, 605 pp. 41–58. 606
13. Miaou, S.P., and H. Lum. Modeling vehicle accidents and highway geometric design 607 relationships. Accident Analysis and Prevention, vol. 25, no. 6, 1993, pp. 689–709. 608
14. Abdel-Aty, M. A., and A. E. Radwan. Modeling traffic accident occurrence and involvement. 609 Accident Analysis & Prevention, Vol. 32, No. 5, 2000, pp. 633-642. 610
15. Maher, M.J., and I. Summersgill. A comprehensive methodology for the fitting of predictive 611 accident models. Accident Analysis and Prevention. Vol. 28, No. 3, 2006, pp. 281–296. 612
16. Miaou, S.P., 1994. The relationship between truck accidents and geometric design of road 613 section: Poisson versus negative binomial regression. Accident Analysis and Prevention Vol. 614 26, No. 4, pp. 471–482 615
17. Srinivas, S., and R. Venkata. Pedestrian crash estimation models for signalized intersections. 616 Accident Analysis and Prevention, Vol. 43, No.1, 2011, pp. 439–446 617
18. Mitra, S., and S. Washington. On the nature of over-dispersion in motor vehicle crash 618 prediction models. Accident Analysis and Prevention Vol. 39, No.3, 2007, pp. 459–468. 619
19. Mitra, S., and S. Washington. On the significance of omitted variables in intersection crash 620 modeling. Accident Analysis & Prevention, Vol. 49, 2012, pp. 439-448. 621
20. Shankar, V., J. Milton., and F. L. Mannering. Modeling accident frequency as zero-altered 622 probability processes: an empirical inquiry. Accid. Anal. Prev. Vol. 29, No. 6, 1997, pp. 829–623 837. 624
21. Anastasopoulos, P.Ch., and F. L. Mannering. A note on modeling vehicle-accident 625 frequencies with random parameter count models. Accident Analysis and Prevention Vol. 41, 626 No.1, 2009, pp. 153–159. 627
22. El-Basyouny, K., and T. Sayed. Accident prediction models with random corridor 628 parameters. Accident Analysis & Prevention Vol. 41, No. 5, 2009, pp. 1118–1123 629
23. Ewing, R., and E. Dumbaugh. The built environment and traffic safety a review of empirical 630 evidence. Journal of Planning Literature, vol. 23, no. 4, 2009, pp. 347-367. 631
24. Miranda-Moreno, L. F., P. Morency., and A. M. El-Geneidy. The link between built 632 environment, pedestrian activity and pedestrian–vehicle collision occurrence at signalized 633 intersections. Accident Analysis & Prevention, Vol. 43, No. 5, 2011, pp. 1624-1634. 634
25. Washington, Simon P., Matthew G. Karlaftis., and F. Mannering., Statistical and Econometric 635 Methods for Transportation Data Analysis, Second Edition; 2011, CRC Press 636
26. Lord, D., and F. Mannering. The Statistical Analysis of Crash-Frequency Data: A Review and 637 Assessment of Methodological Alternatives. Transportation Research Part A, Vol. 44, No. 5, 638 2010, pp. 291–305. 639
27. Lord, D., S. D. Guikema., and S. Geedipally. Application of the Conway Maxwell-Poisson 640 Generalized Linear Model for Analyzing Motor Vehicle Crashes. Accident Analysis and 641 Prevention, Vol. 40, No. 3, 2008, pp. 1123–1134. 642
28. Mannering, F. L., and C. R., Bhat. (2014). Analytic methods in accident research: 643 methodological frontier and future directions. Analytic Methods in Accident Research, 1, 644 2014, pp. 1-22. 645
29. Cameron, A. C., and P. K. Trivedi. Regression-based tests for over dispersion in the Poisson 646 model. Journal of Econometrics. Vol. 46, 1990, pp. 347-364. 647
30. Greene. W. NLOGIT Student Reference Guide Table of Contents. Econometric Software, Inc. 648 2007. 649
31. Anjana, S., and M. V. L. R.Anjaneyulu. Safety analysis of urban signalized intersections 650 under mixed traffic. Journal of safety research, Vol. 52, 2015,pp. 9-14. 651
32. Mohan, D., G. Tiwari.,andS Mukherjee. Urban traffic safety assessment: A case study of six 652 Indian cities. IATSS Research, Vol. 39, No. 2, 2016, 95-101. 653
654
