Master’s Thesis Road Safety and Adverse Weather Conditions An empirical analysis of speeding behavior under wet pavement conditions Submitted at the Faculty of Business, Economics and Social Sciences University of Bern Institute of Business Management and Marketing Section Consumer Behavior Supervisor Dr. Michael Schulte-Mecklenbeck Submitted by Author: Ozan Harman Student ID: 08-926-305 Address: M¨ uhlemattstrasse 59 3007 Bern Place of origin: Zurich, ZH E-Mail: [email protected]Date of Submission: September 13, 2016
56
Embed
Road Safety and Adverse Weather Conditions · PDF fileMaster’s Thesis Road Safety and Adverse Weather Conditions An empirical analysis of speeding behavior under wet pavement conditions
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Master’s Thesis
Road Safety and Adverse Weather ConditionsAn empirical analysis of speeding behavior under wet pavement conditions
Submitted at theFaculty of Business, Economics and Social Sciences
University of Bern
Institute of Business Management and MarketingSection Consumer Behavior
Post-crash first-aid skills cost of vehicle repair rescue facilities
Note: Adapted from Haddon (1972).
combination of different traffic factors and crash phases is shown in Table
1. In addition, exemplary elements of the haddon matrix are presented in
italic letters. The haddon matrix allows a systematic classification of com-
mon accident causes and provides a conceptual framework. In case of an
accident the three factors can be described in three phases. For example, an
alcoholized individual had a tire blowout while driving on a wet road section
(Pre-crash). The driver lost control of his vehicle and crashed into the safety
barriers. Severe injuries were prevented by the airbag and the driver suffered
a minor laceration. However, the driver needed medical care due to bleeding
(Crash). On-site the driver treated the wound with a compression bandage
and received treatment at the hospital in walking distance. Since the vehicle
9
was a limited edition the cost of repairs exceeded the value of the vehicle
(PostCrash).
Even though the haddon matrix has its strength in ad hoc and post hoc
descriptions of traffic situations it has its weakness from a procedural point
of view. It does not identify the driver’s behavioral adaptation to the situa-
tion nor how the accident and the human, vehicle, and environmental factors
affect each other. This is improved in the causal chain model (Elvik, 2004).
To understand this approach we take a step back. Our traffic infrastructure
Figure 1: Causal Chain Model
Note: Reprinted from Elvik (2004).
can be seen as a complex network of numerous road safety measures. To-
days roads are the result of incremental changes to our infrastructure aiming
to provide an efficient system. Traffic incidents are correlated in a way or
another to those road safety measures. As shown in Figure 1 a road safety
measure influences final outcomes through two pathways – target risk factors
and other risk factors. Target risk factors, also called ”engineering effect”,
identify nine factors similar to those classified by the haddon matrix. Their
main characteristic is a high level of physical measurability.
10
According to the causal chain model target risk factors of a road safety
measure can be reduced to one or more of the following types:
1. Kinetic energy2. Friction3. Visibility4. Compatibility5. Complexity6. Predictability7. Road user rationality8. Road user vulnerability9. Forgiveness
Kinetic energy is measured by the mass and speed of a vehicle. Con-
trolled kinetic energy is not a risk as such. However, if the friction between
the tires and the surface is not high enough (skid resistance) the driver will
loose control over the vehicle. Visibility identifies the distance at which a
traffic participant can identify objects. Compatibility refers to differences in
kinetic energy of proximate traffic participants. For example, a tractor is not
allowed on highways due to incompatibility. Another type of risk is the traffic
complexity measured by the amount of information a driver has to process
per unit of time. Closely related to the previous type predictability stands for
the reliability of a driver’s predictions of the immediate traffic situation. For
instance, oncoming traffic on a highway has a high predictability. Individual
rationality denotes the driver’s individual utility maximization. Accordingly,
driver rationality is influenced by multiple aspects such as delay to an ap-
pointment and speeding, alcohol consumption and alertness. Drivers’ are
in this case not minimizing risk but maximizing individual utility. Another
type of risk is a road users vulnerability, e.g., a driver – ceteris paribus –
with a bleeding disorder will behave different, since he is facing higher levels
of injury risk in case of an accident. Finally, system forgiveness measures
11
how forgiving the traffic infrastructure is in case of human and/or technical
failure. For example, safety barriers at steep curves possibly reduce the out-
come severity of an accident.
Up to this point many of the target risk factors can be classified in the haddon
matrix. However the main difference of the two approaches lie in the second
path over which road safety measures influence final outcomes, namely other
risk factors. These risk factors are also called ”behavioral effects” and will
be discussed in the next section.
2.2 Driver Behavior
Road safety measures influence traffic outcomes through behavioral effects.
These effects can be fragmented into the six factors below:
1. How easily a measure is noticed2. Antecedent behavioural adaptation to basic risk factors3. Size of the engineering effect on generic risk factors4. Whether or not a measure reduces injury severity and accident risk5. The likely size of the material damage incurred in an accident6. Whether or not additional utility can be gained
Road users evaluate their environment in order to adapt behavior to
sensory perception. Behavioral adaptation is a result of how a traffic
participant perceives a road safety measure. (1) Road safety measures that
are easily noticed by the driver are more likely to have a behavioral effect.
(2) If a target risk factor has provoked behavioral adaptation in a past
situation a road safety measure affecting this risk factor is more likely to
have a behavioral effect. For instance, the indication of aquaplaning risk
on a highway is more likely to have an effect on drivers’ that are conscious
about the risk of aquaplaning and have adapted their behavior in past
12
situations. (3) Safety measures that have a large engineering effect are more
likely to have a behavioral effect. (4) Accidents are unwanted events and
traffic participants adapt their behavior to road safety measures that affect
the perceived probability and severity of an accident. (5) Drivers’ weight the
consequences of an accident in terms of material damage. Ceteris paribus
an individual will drive more carefully in a parking lot with an expensive
car. (6) Lastly, road users are more likely to change their behavior if this
increases their individual utility.
So far we described theoretical frameworks in which we will discuss our
empirical findings. A missing point is that each individual has different
risk tolerances. In our approach we assume road users to have a invariant
level of risk tolerance. Haight’s (1986) discussion about human risk taking
behavior and the measurement of risk as such seems to complement our
approach rather well. In the following we will present some of his statements
regarding risk. To begin with one has to acknowledge that there is no
universally valid measurement unit such as kilograms or meters to measure
risk. Nevertheless, risk seems to be used synonymous with probability. In
this perspective we have to follow certain rules of probability theory such as
non-additivity of correlated events. Further, risk in the sense of probability
theory requests the measure to be defined in the unit space and to be –
at least in the road safety research – path dependent. Probability theory
is ubiquitous in natural sciences and is helpful to road research from a
technical perspective however formal logic breaks down when individual
risk behavior is studied. Accordingly Shermer (2008) shows that humans
tend to misperceive and miscalculate probabilities. Even though a pure
13
probabilistic approach seems fruitless, driver’s risk believe can be seen as
a mix of probability and outcome expectation. In other words, individuals
subjectively evaluate outcome severity and the probability of an event to
happen. Some attention is required distinguishing subjective and objective
risk. While the former is the risk experienced by the road user, the latter
refers to expert calculations. This distinction is of special interest for
our study analyzing differences in technically required and actual speed
reduction under reduced surface friction.
Individuals somehow balance different risk factors to maximize their indi-
vidual utility with subjective risk as a side condition. Closely linked, the
risk compensation theory assumes that individuals adapt their behavior to
changes in risk factors. For example, improved vehicle safety will lead to
less attentive driving in order to compensate the lower risk level (Wilde,
1976). A rather radical version of risk compensation is represented by the
risk homeostasis theory. It states that no road safety measure will have
any longterm safety improvement effect if the individual risk tolerance
remains unchanged (Wilde, 1988). The risk compensation theory is relevant
to our study in numerous ways. In line with the theory speed (kinetic
energy) should be reduced to compensate for the increased breaking distance
(friction) due to wet pavement.
14
2.3 Rain and Traffic
It is well accepted that precipitation has two main effects. First, crash prob-
ability is higher and second, drivers’ slow down. In the following we will
distinguish between traffic-accident and /-operation research. Prior studies,
analyzing traffic accidents and precipitation, are focusing on crash severity,
crash probabilities, seasonality, amount of rainfall per unit of time, and wet
pavement. There are three prominent methods of analysis dealing with traf-
fic research. 1) The matched pair approach where two measurement periods,
for example a Monday morning, under dry and wet conditions are compared.
2) The difference-in-differences approach where the accident ratios are based
on predicted expectations versus actual accidents in the wet pavement con-
dition rather than a direct comparison of dry and wet conditions. 3) Crash
frequency analysis using count data models.
Closely related to our study, Brodsky & Hakkert (1988) found by using a
difference-in-differences approach that the mean risk of fatal accident was
five (U.S.) to six (Israel) times higher on wet than on a dry pavement. Fur-
ther, in Israel the probability of accidents increases in the period between
November and March, where rain is sporadic. This is in line with our be-
havioral approach suggesting drivers’ risk adaptation to be more prominent
if the individual was previously exposed to the risk factor.
Based on data from a nearby weather station Eisenberg (2004) found a pos-
itive relationship between precipitation and crash rates using a negative bi-
nomial regression. Without exception, literature in which the matched-pair
approach is used present accident risk to be at least 50 % higher under rainy
by a remarkable amount. However, even our smallest sample, Weekends In,
still contains 3’394 observations for both treatment and control group (see
Table 4). The matching procedure allowed to improve balance on all our
samples (for further details see Appendix A).
The balanced samples were then used to plot mean, median, and 95% con-
fidence interval for our dependent variable speed. Since we did not inspect
our samples for normality the confidence intervals were generated using boot-
strapped samples. As shown in Figure 6 our sample confirms that drivers’
slow down under wet pavement conditions. However, there was no signifi-
cant speed difference for weekday evening in and morning out for dry and
wet conditions (see Table 5)1.
1We accept this preliminary parametric testing result without distribution analysis,since we are not interested in the amplitude or position of difference and our samples arelarge.
27
Hence, we assume that there is a non identified covariate inducing this
discrepancy and exclude these observations from our samples.
Note: All measures converted to m/s and m/s2; The speed measure to evaluate the brakingdistance under wet conditions was calculated by subtracting the Pseudomedian from theMdn in the dry state.
As shown in Table 8, the median speed reductions from our results (Pseu-
domedian) are not sufficient if the drivers’ aims to break – ceteris paribus –
within the same distance as in the dry pavement conditions. Based on our
samples we reject our null in favor of the alternative hypothesis (3).
32
5 Conclusion and Discussion
In this study, we have found that a change from dry to wet pavement
conditions lead to a reduction of speed. Across all samples we have found
that speed was reduced by 2.7 to 5.4 percent. For each direction the speed
reduction was higher when average speed was higher. However, the results
are unclear regarding speed dispersion. The dispersion in both directions
for weekends and weekday-middays are homogenous. For weekdays a
change from dry to wet decreases speed dispersion for all subsamples except
weekday mornings in the out direction. Therefore, we cannot make a
general statement regarding speed dispersion. Further, we have found that
drivers’ speed reduction is insufficient to compensate for the risk of lower
pavement friction. In the following we will discuss our assumptions and
findings using Haddon’s matrix (1972) and Elvik’s causal chain model (2003).
5.1 Causal Chain Model and the Haddon Matrix
The haddon matrix helped us to identify variables for our analysis and the
causal chain model to discuss our findings. Beginning with the haddon
matrix, for pre-crash phase human factors we based our analysis assum-
ing drivers’ to differ in their behavior regarding travel purpose (commuting,
leisure and others). We argue that the travel purpose has an impact on the
state of mind of an individual. E.g., commuters might be less attentive be-
cause they are – busy thinking about the oncoming workday (mornings) or
33
tired of the passed workday (evening). We identified travel purpose by dis-
tinguishing between weekends – assuming mostly leisure traffic – and week-
days. Further, we separated mornings, middays, and evenings to identify
commuters in the mornings and evenings. For the crash phase vehicle fac-
tors we assumed that most vehicles are equipped with ABS. This allowed
us to calculate average braking distances under dry and wet conditions to
test our friction risk compensation hypothesis (3). For the pre-crash phase
environmental factors we define two pavement conditions (dry, wet). For
the crash phase environmental factors we consider the route section to be
safe since there is no curvature or crossing traffic. Finally, for the post-crash
phase environmental factors we assume a favorable rescue situation since the
next hospital (Stadtspital Waid) is within 4.5km driving distance. The had-
don matrix is only partially identified since we have no individual level data
about drivers’ or their vehicles’.
Although the haddon matrix is useful to identify variables that potentially
impact traffic, it is inapplicable if we want to explore the relation of cause and
effect between our variables and speed. Therefore, we will asses if our em-
pirical findings make sense from a theoretical point of view using the causal
chain model.
As shown in Table 9 the causal chain model can be used as a checklist for en-
gineering and behavioral effects. The predominant engineering effects regard-
ing wet pavement are friction and visibility. Visibility is impaired through
reflections on the road surface – depending on the characteristics of light
(e.g., sunlight direction, other vehicles’ headlights, and illuminance) and at-
mospheric factors (e.g., fog and rain). Friction is reduced due to lower friction
34
of the pavement surface. Since friction and visibility are reduced by wet pave-
ment, drivers’ should adjust their speed to compensate for the higher risk due
to increased braking distances and impaired visibility. In line with Unrau &
Andrey (2006); Billot et al. (2009) we confirm that drivers’ reduce speed un-
der wet pavement conditions. However, in agreement with Edwards’s (1999)
findings the speed reduction is insufficient for each of our samples regarding
braking distance. Two possible explanations are: 1) Drivers’ misjudge their
driving abilities and the level of risk regarding speed. 2) The total stopping
distance – the sum of the braking distance and the reaction distance – re-
mains unchanged since drivers are more alert under wet pavement conditions
and reduce the reaction distance to compensate the risk from reduced fric-
tion. Road user rationality is adversely affected since the driving situation is
more challenging and the individual utility maximization is prone to human
error. The vehicle is harder to control and the driver depends stronger on
assumptions of the driving situation, his abilities, and the safety equipage of
his vehicle, e.g., rubber composition of the tires, ABS, and other technical
safety measures. Finally, the traffic infrastructure is less forgiving to human
and/or technical failure and the outcome of a collision is likely to be more
severe. Engineering effects by their own do not necessarily induce a change in
driver behavior. To explore how wet pavement affects driver’s behavior a set
of factors need to be assessed. To begin with, drivers’ continuously evaluate
their environment and determine how the level of risk changes. Measures
have to be noticed by the driver to affect their behavior – as wet pavement
can be noticed audibly and visually and drivers’ are sensibilised by their
driving education to the hazards of wet pavement drivers’ are likely to notice
35
surface wetness. Further, drivers’ seem to have a learning curve regarding
risk factors – if an individual faces a risk factor for the first time he is likely
to overlook it. However, if a risk factor such as wet pavement is regularly
noticed the driver learns how to react based on antecedent behavioral adap-
tation. As rain is a frequent event – wet pavement represented 48 percent of
our observations – drivers’ regularly adjust their behavior to wet pavement
conditions. Closely connected to the noticeability the size of an engineering
effect is positively correlated to the change in behavior. Drivers’ adapt their
behavior stronger for great changes in the engineering effect. Two of our
results seem to be affected by the size differences. 1) We found that speed
did not differ for a change from dry to wet pavement for weekday evening
in and morning out observations and assumed that there is some uniden-
tified covariate. We presume that the covariate leading to this bias is the
driver’s domicile. Commuting drivers’ living nearby leave for work in the
morning and come home in the evening. This subpopulation of drivers’ is
overrepresented in our sample predominantly in mornings and evenings lead-
ing to misbalances in our samples. The road section where the study site
is located leads out of the quarter Affoltern (population 25’000). Drivers’
living in Affoltern and frequently using this road section are familiar with the
surroundings and therefore behave differently then other traffic participants.
The perceived complexity and predictability of the driving environment is
lower to those drivers’ living close to this road section and result in a smaller
sized engineering effect. However, the size difference of the engineering effect
is due to the drivers domicile and not the pavement condition. 2) Testing
the homogenous speed dispersion hypothesis (2) we found that dispersion
36
was only affected in weekday mornings and evenings. Therefore, there is no
generally valid answer how dispersion is impacted by wet pavement. This is
in line with previous findings since results are contradictory. While Unrau &
Andrey (2006) found an increase in dispersion, Padget et al. (2001); Liang
et al. (1998) stated dispersion to decrease. Nevertheless, weekday mornings
and evenings are the periods with the highest traffic density and drivers’
are most likely to be commuters. High traffic density and drivers being
commuters increase the complexity of the driving environment. High traffic
density is more demanding since drivers’ have to adjust their behavior to
nearby traffic participants such as oncoming traffic and overtaking vehicles.
Further, commuters are likely to think about their oncoming/passed work-
day making them less attentive. Middays and weekends are less prone to
these biases therefore the size of the engineering effect is smaller. Again, the
size difference of the engineering effect results from other measures than the
pavement condition. Our results are more clear regarding speed as a factor
affecting accident risk, severity, and the size of material damage. Drivers’
seem to acknowledge this relationship in the context of wet pavement. We
found that the relative speed adjustment is higher when the average speed is
higher. This indicates that drivers’ adjustment is overproportional – drivers’
overcompensate the risk of wet pavement and higher speeds. Finally, it seems
reasonable that drivers’ can increase their individual utility by reducing their
speed to compensate for the risk related to wet pavement.
37
Table 9: Causal Chain Model – Effects of Wet Pavement
Engineering effect of wet pavement
Generic risk factors Effect of measure on generic risk fac-tors.
1. Kinetic energy Not affected2. Friction Adversely affected; wet road surfaces
provide less friction.3. Visibility Adversely affected: Shorter visibility
and road surface glare depending onthe level of wetness and illuminance.
4. Compatibility Not affected5. Complexity Not affected6. Predictability Not affected7. Road user rationality Adversely affected: more challeng-
ing driving environment; subjectiveerrors are more likely.
8. Road user vulnerability Not affected9. Forgiveness Adversely affected: infrastructure is
less forgiving; outcomes of humanand technical failure are likely to bemore severe.
Behavioral effect of wet pavement
Factors eliciting behavioral adaptation Effect of measure on factors elicitingbehavioral adaptation.
1. How easily a measure is noticed Wet pavement is easily noticed.2. Antecedent behavioural adaptation
to basic risk factorsDrivers’ are regularly in wet pave-ment situations.
3. Size of the engineering effect ongeneric risk factors
Depends on the situation.
4. Whether or not a measure reducesinjury severity and accident risk
Both injury severity and accidentrisk are increased.
5. The likely size of the material dam-age incurred in an accident
Slight increase.
6. Whether or not additional utilitycan be gained
Utility can be increased by reducingspeed due to lower accident risk.
Note: Adapted from Elvik (2004)
38
5.2 Strengths and weaknesses
By using large samples and robust testing methods we provide rigorous re-
sults for the location Zurich, Switzerland. Further, we used a novel approach
to generate our weather data, which we assume to be both valid and reli-
able. Nevertheless, we based our results on strong assumptions regarding the
sample selection, subsetting, and variables of interest. Our analysis covers a
single study site only with daytime observations for the month of February
2016. Further we acknowledge one should proceed with great caution if con-
clusions about individual level behavior are based on non-individual data.
Nevertheless, by using a paired matching approach we reduce biases.
With respect to variable selection researchers analyzing the effect of adverse
weather use different proxies, measurement units and evaluation methods
when defining weather and speed (see Table 10). Regarding speed data we
assume stationary radar measurements to be superior to those generated
using a handheld radar camera. Accordingly, we expect the handheld mea-
surements to be biased due to the possibility of distance and angle variation.
We acknowledge the limitations of our study considering the myriad possibil-
ities choosing a weather proxy in terms of measurement location (on site or
nearby), evaluation method (sensory or by research personnel), and unit of
measure (precipitation or road surface). However, our approach is favorable
based on following points.
1) We assume on site measurement to be more reliable since the study site
weather conditions are not necessarily identical to those of nearby weather
stations. 2) We believe a driver’s behavior to be strongly influenced by his
39
visual perception.
The closest proxy to a driver’s visual perception is the evaluation of the
on site weather conditions by another individual. Research following this
approach has chosen to base their result on weather data noted on site by
research personnel. Nevertheless, the data generated in this manner is biased
by the personnels perception and is not replicable. We fill this gap by gener-
ating our raw weather data, on site, using surveillance images. The raw data
is then categorized by research personnel into dry, wet, and snow pavement
conditions. Finally, we argue that wet pavement condition is an appropriate
representation of weather because it is the main cause of accidents related to
adverse weather conditions.
Table 10: Research Design
Speed
Handheld radar Stationary radar
Edwards (1999); Padget et al. (2001) Unrau & Andrey (2006); Billot et al. (2009)
Liang et al. (1998)
Weather
On site Nearby weather station
sensory (technical) research personnel
Percipitation Liang et al. (1998) Edwards (1999) Unrau & Andrey (2006); Billot et al. (2009)
Road surface Padget et al. (2001)
40
5.3 Directions for Future Research
For future research, we advise to cover longer study periods at different loca-
tions and traffic speed limits. Second, in order to understand how different
weather factors are interrelated and how they impact traffic behavior we
encourage researchers to introduce further variables such as visibility, illu-
minance and precipitation. Finally, we recommend to collect data on the
individual level to gain a deeper understanding of traffic behavior.