You are on page 1of 18

On the transferability of traffic conflict-based safety

assessment methods: A case of crash frequency-by-severity


prediction models
Presented by:
Md Mazharul (Shimul) Haque

Authors:
Ashutosh Arun and Md Mazharul Haque
Background
• In Australia, 1200 people killed and 39,000
hospitalised annually (Bureau of Infrastructure
Transport and Regional Economics (BITRE),
2020)

• Leading cause of death for children and young


adults aged 5-29 years (World Health
Organisation, 2018)

• Total social cost of road crashes was $29.7 billion


(Australian Automobile Association (AAA), 2017)
≅ 1.3% of Australia’s GDP
Introduction
• Crash data-based analysis is reactive and
unethical

• It is also afflicted by significant statistical


problems like low sample means and small
sample bias due to rare and random nature of
crashes

• Traffic conflicts are more numerous than


crashes and can be readily observed in fields

• Traffic conflicts have found numerous


Safety pyramid of traffic events adapted
applications in traffic engineering
from (Hydén, 1987)
Motivation • Crashes (unobserved extremes) can be estimated
from traffic conflicts (observed extremes) using
Extreme Value Theory (EVT)

• Arun et al. (2021) developed the first methods to


estimate crash frequency by severity from conflict
indicators (MTTC and Delta-V)* using a
multivariate peak-over threshold (POT) method

• Multivariate POT models could be cumbersome to


estimate for each new site; hence transferability of
existing models needs to be investigated

• Transferability analysis also helps determine the


Etiological relationship between traffic events external validity of models
MTTC: Modified Time-to-Collision
Delta-V: Expected post-collision change in a vehicle’s velocity due to a perfectly inelastic collision
Objective
To analyse the transferability of multivariate crash frequency by severity
models

Research Questions

➢ Can the multivariate Peak-over Threshold models be transferred from base


sites to similar target sites?

➢ What is the best approach of transferring base crash frequency by severity


models?
Crash Frequency by Severity Model
Arun et al. (2021) estimated bivariate Logistic Negated MTTC and ∆𝒗 POT Model for predicting total, severe (MAIS3+),
and non-severe rear-end crashes at two signalised intersections in Brisbane, Australia.
Parameter Values
Marginals 𝝈
ෞ𝟏 (SE) 0.141 (0.008)

𝝃𝟏 (SE) -0.078 (0.043)
𝒖𝟏 -0.51
ME1 821
𝝈
ෞ𝟐 (SE) 1.811 (0.098)

𝝃𝟏 (SE) 0.061 (0.043)
𝒖𝟐 11.16
ME2 637
Dependence 𝝋 0.999 (2x10-6)
JE 72
Estimated 𝑵𝒕𝒐𝒕𝒂𝒍 5.073
Estimated 𝑵𝒔𝒆𝒗𝒆𝒓𝒆 0.417
Estimated 𝑵𝒏𝒐𝒏−𝒔𝒆𝒗𝒆𝒓𝒆 4.656
Legend:
𝑁𝑡𝑜𝑡𝑎𝑙, 𝑁𝑠𝑒𝑣𝑒𝑟𝑒 , 𝑎𝑛𝑑 𝑁𝑛𝑜𝑛−𝑠𝑒𝑣𝑒𝑟𝑒 : Number of the annual total, severe, and non-severe crashes, respectively
𝝈𝟏 , 𝝃෢𝟏 , 𝒖𝟏 ) 𝒂𝒏𝒅 (ෞ
(ෞ 𝝈𝟐 , 𝝃෢𝟐 , 𝒖2 ) are the scale, shape, and location parameters for the negated MTTC and ∆𝑣
margins, respectively
ME1 and ME2: Marginal Exceedances of negated MTTC and ∆𝑣 margins, respectively; JE: Joint Exceedances;
SE: Standard Error; 95% CI: Confidence Interval at 95% level of significance

MAIS3+: Maximum Abbreviated Injury Scale ≥ 3


Study Design
Uncalibrated Approach
• Fundamental property of extreme value models wherein the number of conflict
threshold exceedances is directly proportional to the number of expected
crashes
𝑁𝑐𝑟𝑎𝑠ℎ𝑒𝑠,𝑏𝑎𝑠𝑒 𝑚𝑜𝑑𝑒𝑙
• 𝑁𝑐𝑟𝑎𝑠ℎ𝑒𝑠,𝑡𝑎𝑟𝑔𝑒𝑡 = 𝑁 × 𝑁𝑒𝑥𝑐𝑒𝑒𝑑𝑎𝑛𝑐𝑒𝑠,𝑡𝑎𝑟𝑔𝑒𝑡
𝑒𝑥𝑐𝑒𝑒𝑑𝑎𝑛𝑐𝑒𝑠,𝑏𝑎𝑠𝑒 𝑚𝑜𝑑𝑒𝑙

• For estimating total crashes at a target intersection, the marginal exceedances


of MTTC (ME1) were 𝑁𝑒𝑥𝑐𝑒𝑒𝑑𝑎𝑛𝑐𝑒𝑠,𝑚𝑜𝑑𝑒𝑙, the number of MTTC values in the
target dataset exceeding MTTC threshold (𝒖𝟏) was 𝑁𝑒𝑥𝑐𝑒𝑒𝑑𝑎𝑛𝑐𝑒𝑠,𝑡𝑎𝑟𝑔𝑒𝑡

• Similarly, frequency of severe crashes was estimated; non-severe crashes was


the difference between total and severe crashes
Threshold Calibration Approach
• Conflict thresholds of both negated MTTC and ∆𝑣 margins were calibrated
using target site data

• Usual threshold estimation methods for extreme value models, namely,


Mean Residual Life Plot, Threshold Stability Plot, and Spectral Measure
Plots were used
Crash Frequency by Severity
Estimation
𝑇 𝑇
• 𝑁𝑇,𝑡𝑜𝑡𝑎𝑙 = × Pr −𝑀𝑇𝑇𝐶 ≥ 0 = × 1 − 𝐹𝑀𝑇𝑇𝐶 0
𝜏 𝜏

𝑇 𝑇
• 𝑁𝑇,𝑠𝑒𝑣𝑒𝑟𝑒 = × Pr −𝑀𝑇𝑇𝐶 ≥ 0 ‫ ≥ 𝑣∆ ٿ‬16 = × 1 − 𝐹𝑀𝑇𝑇𝐶,∆𝑣 0,16
𝜏 𝜏

• 𝑁𝑇,𝑛𝑜𝑛−𝑠𝑒𝑣𝑒𝑟𝑒 = 𝑁𝑇,𝑡𝑜𝑡𝑎𝑙 − 𝑁𝑇,𝑠𝑒𝑣𝑒𝑟𝑒


where, 𝑁𝑇,𝑡𝑜𝑡𝑎𝑙 and 𝑁𝑇,𝑠𝑒𝑣𝑒𝑟𝑒 are the number of total and severe crashes, respectively;
𝜏 is the conflict observation period (in hours);
T is the desired crash estimation period, equal to 1 year (=365×24 hours);
The crash thresholds for the MTTC and ∆v indicators, namely, 0 s and 16 m/s, respectively.
• The ∆v threshold was derived from the ∆v-MAIS3+ injury relationship developed by Bahouth et al.
(2014)
Conflict Data
Collection
Intersection Camera Camera Data collection schedule
# position
Stafford Rd – 1 View 1 12/11/2019 (Tues; 6 am to 6 pm),
Appleby Rd 13/11/2019 (Wed; 6 am to 6 pm)
(SA) View 2 14/11/2019 (Thurs; 6 am to 6 pm),
Intersection 15/11/2019 (Friday; 6 am to 6 pm)
2 View 1 12/11/2019 (Tues; 6 am to 6 pm),
13/11/2019 (Wed; 6 am to 6 pm)
View 2 14/11/2019 (Thurs; 6 am to 6 pm),
15/11/2019 (Friday; 6 am to 6 pm)
Gold Coast Hwy – 1 View 1 12/11/2019 (Tues; 6 am to 6 pm),
Hope Island Rd 13/11/2019 (Wed; 6 am to 6 pm)
(GH) Intersection View 2 14/11/2019 (Thurs; 6 am to 6 pm),
15/11/2019 (Friday; 6 am to 6 pm)
2 View 1 12/11/2019 (Tues; 6 am to 6 pm),
13/11/2019 (Wed; 6 am to 6 pm)
View 2 14/11/2019 (Thurs; 6 am to 6 pm),
15/11/2019 (Friday; 6 am to 6 pm)
Conflict Data Extraction
Examples of Identified Conflicts
Intersection-wise descriptive
statistics of crash and conflict data
Intersection No. of total No. of severe Conflict Mean Median Std. dev.
name rear-end rear-end indicator
crashes crashes (units)
(2015-19) (2015-19)
Stafford Rd – 6 0 MTTC (s) 1.01 0.88 0.53
Appleby Rd –
St (SA) ∆𝒗 (m/s) 6.31 6.17 3.16
Intersection
Gold Coast 12 0 MTTC (s) 0.85 0.86 0.23
Hwy – Hope ∆𝒗 (m/s) 5.33 5.01 2.13
Island Rd
(GH)
Intersection
Results
Gold coast hwy-Hope Island Rd Intersection Stafford Rd-Appleby Rd Intersection

Mean Crash Frequency


Mean Crash Frequency

Observed Uncalibrated Threshold Complete


Observed Uncalibrated Threshold Complete
Calibration Re-estimation
Calibration Re-estimation
Model Transfer Approach Model Transfer Approach
Conclusions
• Crash frequency-by-severity methodology is externally valid
➢ Uncalibrated approach yields accurate enough crash estimates
• Threshold calibration approach is the most optimal model transfer approach
➢ Uncalibrated approach might not be reliable for all target sites
➢ For target sites within similar jurisdictions as base sites, simple calibration of conflict
thresholds results in improvements in the accuracy of crash estimates
➢ For target sites in other jurisdictions, calibration of scale, shape, and dependence
parameters may be required
• Complete re-estimation of models suffers from the weakness of intersection-
wise models
➢ This approach cannot take advantage of the sharing of conflict information among
sites that is present in combined dataset (base) models
References
• Arun, A., Haque, M. M., Bhaskar, A., Washington, S., Sayed, T., 2021. A bivariate extreme value model for estimating crash frequency by
severity using traffic conflicts. Analytic Methods in Accident Research, 32, 100180.

• Australian Automobile Association (AAA), 2017. Cost of road trauma in Australia. Canberra ACT: Australian Automobile Association (AAA).

• Bureau of Infrastructure Transport and Regional Economics (BITRE), 2020. Road trauma Australia 2019 statistical summary. Canberra ACT:
BITRE.

• Hydén, C., 1987. The development of a method for traffic safety evaluation: The Swedish Traffic-Conflicts Technique. International Calibration
Study of Traffic Conflict Techniques.

• World Health Organisation, 2018. Global Status Report on Road Safety 2018. Geneva: World Health Organisation.
Suggested Readings
• Arun, A., Haque, M. M., Bhaskar, A., Washington, S., Sayed, T., 2021b. A systematic mapping review of surrogate
safety assessment using traffic conflict techniques. Accident Analysis and Prevention, 153, 106016.

• Arun, A., Haque, M. M., Washington, S., Sayed, T., Mannering, F., 2021c. A systematic review of traffic conflict-
based safety measures with a focus on application context. Analytic Methods in Accident Research, 32, 100185.

• Arun, A., Haque, M. M., Bhaskar, A., Washington, S., Sayed, T., 2021a. “How Many are Enough?” Investigating the
effectiveness of multiple conflict indicators for crash frequency-by-severity estimation through automated traffic
conflict analysis. under review at Transportation Research Part C: Emerging Technologies.

• Zheng, L., Sayed, T., Essa, M., 2021. Investigating the transferability of Bayesian hierarchical extreme value model
for traffic conflict-based crash estimation. Canadian Journal of Civil Engineering, 48 (9), 1071-1080.

You might also like