Contents lists available at ScienceDirect
Accident Analysis and Prevention
journal homepage: www.elsevier.com/locate/aap
Modelling of road traffic fatalities in India
Rahul Goel
MRC Epidemiology Unit, University of Cambridge, UK
A R T I C L E I N F O
Keywords:
India
Traffic fatalities
Distance-decay functions
Accident prediction model
Ecological model
A B S T R A C T
Passenger modes in India include walking, cycling, buses, trains, intermediate public transport modes (IPT) such
as three-wheeled auto rickshaws or tuk-tuks, motorised two-wheelers (2W) as well as cars. However, epide-
miological studies of traffic crashes in India have been limited in their approach to account for the exposure of
these road users. In 2011, for the first time, census in India reported travel distance and mode of travel for
workers. A Poisson-lognormal mixture regression model is developed at the state level to explore the relationship
of road deaths of all the road users with commute travel distance by different on-road modes. The model
controlled for diesel consumption (proxy for freight traffic), length of national highways, proportion of popu-
lation in urban areas, and built-up population density. The results show that walking, cycling and, interestingly,
IPT are associated with lower risk of road deaths, while 2W, car and bus are associated with higher risk.
Promotion of IPT has twofold benefits of increasing safety as well as providing a sustainable mode of transport.
The mode shift scenarios show that, for similar mode shift across the states, the resulting trends in road deaths
are highly dependent on the baseline mode shares. The most worrying trend is the steep growth of death burden
resulting from mode shift of walking and cycling to 2W. While the paper illustrates a limited set of mode shift
scenarios involving two modes at a time, the model can be applied to assess safety impacts resulting from a more
complex set of scenarios.
1. Introduction
India has one of the highest shares of road traffic fatalities in the
world. A large proportion of these fatalities are pedestrians, cyclists,
and riders of motorised two-wheelers (2W) (Hsiao et al., 2013; Mohan
et al., 2015). This is because a large share of daily trips is contributed by
the three modes. According to Census 2011 in India (Census-India,
2017a), the three modes contribute up to 70% of work trips. India lacks
government-led efforts for transport-related data in terms of travel
surveys or traffic counts. As a result, road traffic injury models been
limited in their approach to account for exposure of multiple road user
groups. At most, models have used vehicle registration numbers of 2W
and cars which highly overestimate actual in-use fleet (Goel et al.,
2015; 2016). Moreover, registration data does not account for walking,
cycling and use of public transport (PT) and is therefore limited in its
application.
Current levels of vehicle ownership in India are far lower than most
high-income countries. In 2011, only 6% of all the households in India
owned a car, compared to more than 75% households in many high-
income countries (Census-India, 2017b; Statista, 2017). As a result, a
large proportion of population continues to walk, cycle, or use PT. At
the same time private vehicle ownership witnesses an inevitable
growth. From 1990 to 2015, the average year-on-year growth rate of
2W and cars was 9–10%, implying that private motorised fleet is dou-
bling every 7–8 years. This rate is many times higher than the growth
rate of population and, therefore, indicates a dramatic mode shift from
walking, cycling and PT to private vehicle use.
Motorisation in India is also different from many of the high-income
countries in two main aspects. Firstly, motorised traffic in India is
dominated by 2W. For every car in India, there are more than five times
as many 2W (MoRTH, 2013). In case of a crash, ceteris paribus, a 2W
rider is many times more vulnerable to an injury than a car driver.
Thus, a motorisation based on 2W makes its road users more risk-prone.
A car-based motorisation, on the other hand, ensures higher safety of
vehicle occupants. Secondly, PT modes in India include not only buses
and trains but also a range of other intermediate modes such as auto
rickshaws and tuk-tuks, common in many south-Asian settings. They
serve the purpose of PT and, at the same time, have a smaller engine
capacity than a bus or even a car. Thus, an impact of these modes on
safety is important from the perspective of transport policies.
In summary, travel patterns in India present a complex and a unique
mix of traffic modes and are going through rapid changes. All these
changes are occurring in the context of poor enforcement of traffic laws,
as well as a lack of safe infrastructure for walking and cycling. Given
this background, travel patterns are likely to be a strong predictor of the
number of road injuries. This is the first time Census in India has
https://doi.org/10.1016/j.aap.2017.12.019
Received 29 October 2017; Received in revised form 25 December 2017; Accepted 27 December 2017
E-mail addresses: rg574@medschl.cam.ac.uk, rahulatiitd@gmail.com.
Accident Analysis and Prevention 112 (2018) 105–115
0001-4575/ © 2018 The Author. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/BY/4.0/).
T
included travel-related information. The information includes the mode
of travel and the travel distance for workers. This gives an opportunity
to explore how the travel patterns are related to road deaths to assess
and design travel demand and traffic safety policies in India.
2. Objectives
The main objective of this study is to develop an ecological model of
road traffic fatalities, with states of India as areal units. The model aims
to establish a relationship between total annual road fatalities and
commute travel distance by different modes, while controlling for state-
specific confounders. I aim to develop a model with a form often used in
injury modelling and shown in Eq. (1):
= ∑n M M M e1 2 3e e e β x1 2 3 i i (1)
where, M1, M2, and M3 represent travel distance (or volume) of the
three road user categories, e1, e2, and e3 represent their respective
exponents, xi represents a set of predictor variables which control for
factors other than volume, and βi their corresponding coefficients. The
values of exponents and coefficients are obtained using regression
modelling.
The three road user categories have been used only for illustration.
This form of the model is achieved by anti-logging a log-linear re-
lationship between injury counts (n) and the volume or distance vari-
ables (M1, M2, M3). It is a usual practice to include these variables in
their logged form. This also results in multiplicative risk factors as
shown in Eq. (1). Such models are often referred to as accident pre-
diction models. This name, however, is an oxymoron, since accidents by
nature are not predictable, and therefore a more scientific term should
be injury prediction models.
The models have been developed at a range of level of aggregation.
These includes traffic junctions, roundabouts, crossings, or road sec-
tions among the ‘micro ‘or ‘meso-level’ models to city wards, traffic
zones, and municipalities among the ‘macro-level’ or ecological models
(Elvik and Bjørnskau, 2017). For this paper, I will only discuss models
at areal or macro levels. The outcome variable in these models also vary
based on the objective of the study. Most models include number of
injuries or crashes of a specific road user as outcome. In such models,
the exposure variables include the volume of that road user (injuries of
which are outcome variable) along with the volume of conflicting road
user. For instance, models with pedestrian injuries as outcome and
pedestrian and car volume as explanatory variables. No model in the
literature has accounted for more than two road users, except Elvik
(2016) who modelled pedestrian injuries using volume of cars, cyclists
and pedestrians.
The model presented in this paper differs from the previous litera-
ture in two main aspects. First, the dependent variable in the model is
the number of road deaths of all road users and not specific to a single
road user. Second, the model accounts for multiple modes as ex-
planatory variables, and not just two modes, thus reflecting the het-
erogeneity of traffic on Indian roads. Thus, the model in this paper aims
to establish a relationship between overall road death burden and a mix
of travel modes. This also implies that a comparison of the results
presented in this paper with the literature needs to be done cautiously.
My aim to develop this model is twofold—analytical and for pre-
diction. The former will be achieved my assessing the magnitude and
signs of exponents of different road users. The latter will be achieved by
simulating future travel patterns to assess their impact on road deaths. I
will explain these using an example. Suppose that there are three road
users in the model specified in Eq. (1), and from the regression mod-
elling it is estimated that the two of them (say, M1 and M2) have po-
sitive exponents (e1 and e2) and one (say, M3) has a negative exponent
(e3).
From an analytical perspective, this implies that an increase in M1
and M2 will increase injury burden, while an increase in M3 will reduce
it. Among M1 and M2, the comparison between the magnitudes of their
exponents will also illustrate which of the two modes will result in
higher injury burden if both are increased by the same amount. From a
prediction perspective, one can model what-if scenarios of mode shift
and understand the trajectories of road death burden. For instance,
mode shift from M3 (mode associated with less risk) to M1 or M2
(modes associated with higher risk) will result in much higher death
burden than mode shift within M1 and M2.
The literature on accident prediction models is also divided among
those where the exposure of different road users (such as M1, M2, and
M3 in the example above) are in the form of counts (or volumes) and
those where it is in the form of distance (Elvik and Bjørnskau, 2017;
Schepers and Heinen, 2013). When the units of analysis are point lo-
cations or of a consistent size, the counts can be justified as an exposure
variable. For instance, counts of motor vehicles, pedestrians, or cyclists
at traffic junctions, road sections, or traffic analysis zones in a city. If
the units of analyses differ in their size, the counts may be an in-
complete measure. The models with only counts also eliminate the
possibility to predict changes in injuries if population travelled using
the same modes however the distance of travel changed. Therefore, a
model with distance is more robust in its application to predict changes
in injuries resulting from changing travel patterns.
3. Data
The model explained in the previous section needs three main data
types—a) annual number of road deaths for each state as dependent
variable, b) mode-specific commute travel distance, and c) other ex-
planatory variables. In 2011, India had 28 states and 7 Union
Territories (UTs). The average population of the UTs is 2.9 million
while that of the states is 41 million. Two of the UTs are islands,
Andaman and Nicobar Island in the east and Lakshadweep in the west,
and contribute 0.04% of the total population of the country. These were
excluded from the analysis. The remaining 28 states and 5 UTs will be
referred to as 33 states henceforth. Note that Delhi, the capital city of
India, is a city-state and is therefore included as one of the units in this
analysis. The states cover a large range of population from 0.24 million
to 200 million. Table 1 presents the descriptive statistics.
National Crime Records Bureau of India publishes annual number of
road accidents, number of people injured, and number of deaths for
each state and UT in India. Number of injury crashes are highly un-
derestimated in India (Mohan et al., 2015), and as a result only number
of deaths have been used for the analysis. Corresponding to census year,
I used average number of fatalities for the three years (2010 through
2012) for stable estimates (NCRB, 2011; 2012; 2013).
In almost all the states, year-to-year variation of number of road
deaths was minimal across the three years (see Appendix: Table A1).
There are, however, two exceptions—Punjab, where number of road
deaths corresponding to three years are 2133 (2010), 4897 (2011) and
4795 (2012), and Nagaland, with 71 (2010), 106 (2011), and 44 (2012)
deaths. In both the states, highest number of deaths is more than 2
times higher than the lowest number. Average fatality rate across the
states is 11.6 per 100,000 persons and vary from 2.3 to 22.4. Fig. 1
presents fatality rates for all the 33 states and overall India in a des-
cending order.
In this analysis, I have excluded deaths occurring at railway cross-
ings, which is an area where on-road modes and trains interact. Over
the three years, total number of road deaths at railway crossing are
3344 (2010), 2366 (2011), and 1808 (2012). In contrast, total number
of road deaths for the three years are 133938, 136834 and 139091
respectively (NCRB, 2011; 2012; 2013), thus, deaths on railway
crossing is 1–3% of the on-road deaths.
In 2011, Census of India introduced two questions regarding the
commute of workers (Census-India, 2017a). These questions were asked
from a subset of all workers—the category called ‘other workers’. This
category excludes those involved in agricultural or household-based
activities. The category of ‘other workers’ represent 42% of all the
R. Goel Accident Analysis and Prevention 112 (2018) 105–115
106
workers in India (Census-India, 2017c). The two questions on com-
muting included mode of travel and one-way distance (in kilometres)
from residence to place of work, and in the former only one mode could
be selected. For further details see Goel, 2018.
The question on mode thus disregards the multimodal character-
istics of some of the trips. However, census provides no details in this
regard (Census-India, 2017d). Thus, the working assumption is that the
respondents informed their main mode of travel—the one using which
they covered the longest travel distance. Since the census is conducted
using personal interviews, it is possible that these questions, in some
cases, were answered by proxy respondents, for instance, by other
members of the household. However, no such information is available
from census to account for this bias.
There are 9 options for the travel modes: (1) walk, (2) cycle, (3)
moped/scooter/motorcycle, (4) car, (5) tempo/auto rickshaw/taxi, (6)
bus, (7) train, (8) water transport, and (9) any other, and an option of
‘No travel’. Category 3 is referred to as motorised two wheelers (2W),
and category 5 as IPT. The latter consists of intermediate public
transport, or para-transit modes such as three-wheeled auto rickshaws,
common across India (for their description see Goel and Tiwari, 2016
and Kumar et al., 2016).
For each mode, Census has reported mode-specific count of workers
classified into 7 distance categories: 0–1 km, 2–5 km, 6–10 km,
11–20 km, 21–30 km, 31–50 km, and>50 km. Walking has been re-
ported up to 10 km, and cycling up to 30 km. The data has been re-
ported only at the aggregate level of states and districts, with a further
classification into rural, urban and total. In this analysis, total data
(urban plus rural) has been used at state level. Also, water transport and
‘any other’ categories were excluded. These modes were reported by
1.2% of those travelling by one of the 9 travel modes. The detailed
method to estimate average distance travelled by each mode in each
state is presented in Goel (2018). I used average distance (see Ap-
pendix: Table A2) and multiplied by corresponding number of workers
to estimate total commute distance travelled by each mode. Table 1
presents the descriptive statistics of the total distance.
One of the major limitations of distance estimates from census data
is that it is limited to only commuting while our dependent variable
includes road deaths from all travel. Therefore, I investigate the re-
lationship between commute distance and the variables which re-
present overall travel. A study was commissioned by MoPNG to esti-
mate sectoral share of petrol and diesel consumption in India.
According to this, Petrol is only used by cars and 2W, except 2% by IPT
(Nielsen, 2013). Thus, petrol consumption is a good indicator of total
distance travelled by 2W and cars. The Pearson correlation of annual
petrol consumption at state level (MoPNG, 2012) with commute dis-
tance estimated for 2W is 0.98 and with commute distance of car is
0.92. This indicates that commute distance by these two modes is a
good proxy of overall travel distance by the two modes. Distance by car
has less correlation than 2W since cars also use diesel and, to a lesser
extent, compressed natural gas (MoPNG, 2012; Goel and Guttikunda,
2015).
In case of buses, I use distance travelled by public buses to compare.
These buses are operated by government-run organisations known as
State Road Transport Undertakings (SRTUs). There is a correlation of
0.89 between passenger kilometres reported by SRTU buses in the states
(MoRTH, 2011) and the commute distance travelled by buses. The data
for SRTU was available for 17 out of 33 states. Note that bus transport
in India is carried out by public as well as private buses. Therefore,
public buses do not represent all the buses, yet they represent bus travel
for all purposes.
Table 1
Descriptive statistics of variables.
Mean Median Std. Deviation Minimum Maximum
Average annual number of fatalities (2010–2012) 4139 2221 4894 24 15669
Population 36,679,089 25,351,462 44,957,638 243,247 199,812,341
Average fatality rate (per 100,000 population) 11.6 12.1 4.9 2.3 22.4
Total commute distance by Walk (km) 2,817,833 1,874,778 3,060,533 39,947 11,049,986
Total commute distance by Cycle (km) 4,318,576 2,844,264 5,808,973 3738 26,810,567
Total commute distance by 2W (km) 6,064,611 3,876,805 7,491,930 13,554 28,539,315
Total commute distance by Car (km) 2,342,221 1,163,472 2,716,398 13,581 9,114,740
Total commute distance by IPT (km) 1,729,288 917,165 2,415,759 25,242 8,979,712
Total commute distance by Bus (km) 14,271,637 5,772,699 18,475,298 49,906 82,321,109
Total commute distance by Train (km) 9,597,981 2,468,720 18,804,973 8212 88,281,102
Annual Diesel Consumption (×1000 tonnes) 1957 930 2217 48 7483
Percent Urban Population 38% 30% 22% 10% 98%
Length of national highway (km) 2009 1512 1790 1 5874
Built-up Density (persons per km2) 12260 11062 6347 2582 26066
0
5
10
15
20
G
oa
Ta
m
il 
Na
du
Pu
du
ch
er
ry
Ha
ry
an
a
An
dh
ra
 P
ra
de
sh
Da
dr
a 
&
 N
ag
ar
 H
av
el
i
Hi
m
ac
ha
l P
ra
de
sh
Ka
rn
at
ak
a
Pu
nj
ab
Ra
ja
st
ha
n
Gu
ja
ra
t
Ch
ha
ƫ
sg
ar
h
M
ah
ar
as
ht
ra
Ke
ra
la
Ch
an
di
ga
rh
De
lh
i
Si
kk
im
M
ad
hy
a 
Pr
ad
es
h
In
di
a
Da
m
an
 &
 D
iu
Ar
un
ac
ha
l P
ra
de
sh
Ja
m
m
u 
&
 K
as
hm
ir
O
di
sh
a
UƩ
ar
ak
ha
nd
UƩ
ar
 P
ra
de
sh
M
izo
ra
m
As
sa
m
M
eg
ha
la
ya
Tr
ip
ur
a
Jh
ar
kh
an
d
W
es
t B
en
ga
l
M
an
ip
ur
Bi
ha
r
Na
ga
la
nd
(d
ea
th
s p
er
 1
00
,0
00
 p
er
so
ns
)
Fig. 1. Average annual road fatality rates across states and all India.
R. Goel Accident Analysis and Prevention 112 (2018) 105–115
107
There is no data to see correlations for walk, cycle, IPT or trains.
However, given the high correlations for other modes, it can be as-
sumed that these correlations will be consistent across all the modes.
Commute pattern as a proxy for overall travel has also been reported for
England. Goodman (2013) reported strong linear relationships between
mode shares from Census-reported commuting and those from national
travel survey (which includes all trips) for multiple transport modes in
England.
To control for state-specific factors other variables need to be in-
cluded. While distance travelled by passenger transport modes are ac-
counted for, a large proportion of road deaths are caused in crashes
with goods vehicles (Goel, 2017; Mohan et al., 2016; Naqvi and Tiwari,
2017). According to a study commissioned by the Ministry of Petroleum
and Natural Gas (MoPNG), on an average, 70% of the diesel in the
country is consumed by road transport vehicles in India. Within road
transport (70%), 22% is consumed by cars, 28% by trucks, 10% by
buses, and 6% by three-wheeled passengers and goods vehicles
(Nielsen, 2013). Taxis, included within cars, run mostly on diesel (Goel
and Guttikunda, 2015). Thus, diesel consumption can be used as a
proxy of vehicle kilometres travelled by freight as well as taxis. Annual
diesel consumption for year 2011–12 was used from annual publication
of MoPNG (MoPNG, 2012).
The other variables include proportion of population living in urban
areas, population density and length of national highways. All the three
variables can be categorised as built environment variables and are
expected to interact with the exposure of road users in positive or ne-
gative direction. Urban areas have different traffic patterns than rural
areas, population density have been reported to have a significant effect
on traffic safety and length of national highway determines inter-state
connectivity and amount of long-distance traffic.
Census reports population classified by urban and rural. I calculated
the proportion of total population living in urban areas and refer to it as
level of urbanisation. The level of urbanisation varies from 10% to 98%
(Census-India, 2012). Population density of the states were calculated
using state-level urban built-up area reported by the National Remote
Sensing Centre through their web portal, Bhuvan (NRSC, 2016). The
NRSC reports Land use Land Cover data using multi-temporal satellite
data of 2011–12 from Resourcesat-2 LISS III images. We used the sum of
urban and rural areas reported in NRSC data and expressed the density
as persons per km2. Only national highways have been used since their
reporting is likely to be more reliable and consistent as they are
maintained by a single organisation across the country—National
Highway Authority of India. The length of NHs reported for year 2011
were used (MoRTH, 2013).
4. Method
4.1. Cluster analysis
The regression model is aimed at understanding the independent
effects of distance travelled by different modes. However, it is also
worth understanding how travel patterns vary across the states, and
how this variation relates to fatality rates. Within each state, I expressed
total commute distance (estimated as described in previous section) of
each mode as the proportion of total commute distance across all the
modes. The proportions for each state total to unity and I refer to these
as mode shares. Given a large variation in the magnitude of total dis-
tance travelled, mode shares enable comparison across the states.
Following this, I used k-means clustering to classify states into a
group of clusters with similar distribution of mode share. In this
method, states are assigned into clusters such that the sum of the
squared deviations (Euclidian distance) from each observation and the
cluster centroid is minimised. This sum is called within-cluster sum of
squares (WSS). States were classified into 5 clusters, based on the op-
timum value of total WSS across all clusters. This means that an in-
crement in the number of clusters beyond 5 does not further reduce the
value of WSS.
4.2. Regression model
As this analysis is at the state level, it is constrained by a small
number of observations (n=33). In this context, use of Bayesian
methods rather than frequentist method is preferred especially if var-
ious explanatory variables are included (Hox et al., 2012). I modelled
fatalities with Poisson-lognormal mixture using Bayesian hierarchical
modelling. The modelling was done using R-INLA (Rue et al., 2009)
which is an R package and employs integrated nested Laplace approx-
imations to estimate the posterior distributions. The package has been
used for injury modelling by DiMaggio (2015) and Goel et al. (2018).
The hierarchical model is described as follows:
=y Poisson f( )n n (2)
= + + +f elog( ) log( ) β βX δn n n n0 (3)
∼ N τδ (0,1/ )n δ (4)
∼τ logGammalog( ) (1,0.0005)δ (5)
where, ynare the observed annual fatality counts of all road users in
state n, fn are the expected count of fatalities, Xn represents a vector of
explanatory variables, en is the exposure, β0 is the intercept, β is a
vector of fixed effect parameters, and δn is the uncorrelated hetero-
geneity or unstructured error. Here, δn represents overdispersion and
accounts for variation in the expected fatality risk after controlling for
the independent variables. Population of each state represents the
corresponding exposure (e )n .
The first level of the hierarchical modelling framework presented in
the Eq.(2) through (5) is the likelihood model or the random sampling
of number of fatalities (y )n from a Poisson distribution with a state-
specific expected count ( fn). The second level models the log-linear
relationship between expected fatality risk and independent variables.
The logged relationship ensures positive values of f .n Note that exposure
(ei) is an offset (a covariate with coefficient value 1) and, therefore,
effectively acts as a denominator for left-hand side of the equation and
expresses it as population risk (log (λ )n = f elog( / )n n ). Therefore, this
modelling framework accounts for exposed population explicitly, rather
than treating it as a covariate.
The prior for δn is modelled as normal distribution where τδ refers to
the precision of the distribution and is inverse of the variance. Note that
from Eq. (3), fn can be expressed as e e e. .β βX δn n0 , therefore, in effect, the
log of the overdispersion term (eδn) has a normal distribution, thus the
distribution of eδnis lognormal. This makes the model a Poisson-log-
normal mixture model. Further, the hyper parameter, τlog( δ) is assigned
a prior of log-gamma distribution with shape and inverse-scale para-
meters of 1 and 0.0005, respectively. Using log of τδ ensures a positive
value as it represents standard deviation. These priors with large
standard deviations assume no prior knowledge of the effects and are
referred to as uninformative priors.
Census only reports main mode of travel, therefore, in case of PT,
this will result in underestimation of walking to and from the PT stops.
To account for this, I assumed 1 km of walking distance corresponding
to each trip of a PT mode (IPT, bus and train) longer than 1 km. The
sensitivity of this assumption on the model results was tested by as-
suming no walking distance and assuming 1.5 km as the walking dis-
tance. While train distance is not included in the model, its walking
distance is included as this distance is covered on the roads. Note that
1 km includes walking on either side of a PT trip, and is likely a con-
servative estimate for many trips which involve much longer than
500m to walk to PT stops.
Four regression models are presented varying from minimally con-
trolled to maximally controlled. The additional explanatory variables
are added into the models in three stages. Model 1 (minimally con-
trolled) includes only commute distance variables, model 2 adds diesel
R. Goel Accident Analysis and Prevention 112 (2018) 105–115
108
consumption, model 3 has additional adjustment of length of National
Highways (model 3), andmodel 4 also includes population density and
percent urban population. Model 1 accounts only for passenger travel
modes, model 2 also accounts for freight modes, and models 3 and 4
account for built environment and related features.
The estimates of Bayesian modelling are in the form of posterior
distributions of all parameters—coefficients as well as error terms. The
results of these distributions will be presented in the form of mean and
standard deviation, after ensuring that the distributions are normally
distributed in which case mean is a suitable central tendency. Their
significance will be reported based on whether zero lies within 90% or
95% Bayesian Confidence intervals (BCI). A coefficient significant at
95% BCI, for example, means that the range which covers 95% of all the
values of a posterior distribution does not include zero.
5. Results
5.1. Cluster analysis
The result of cluster analysis is presented in Fig. 2 which also in-
cludes all India. The figure includes mode shares, fatality rates and the
cluster number. The cells for mode shares and fatality rates are
highlighted with colour scales. The shades of red represent values
greater than the average across all the states, white represents average
values, and shades of green represent values lower than average. This
can also be used to identify modes that define a cluster. I titled each
cluster with a group of modes that the cluster comprises at their above-
average levels represented with shades of red.
Table 2 presents average mode shares within each cluster. Within
each cluster, the mode with a high share is also highlighted, same as
those in the cluster titles in Fig. 2. Cluster 1 consists of states with high
levels of walk, IPT and car, cluster 2 consists of states with high level of
bus use, cluster 3 consists of high levels of cycle and train, cluster 4 has
State Walk Cycle IPT 2W Bus Car Train Fatality Rate Cluster
Arunachal Pradesh 0.27 0.05 0.04 0.14 0.21 0.26 0.02 9.7 1
Meghalaya 0.13 0.01 0.12 0.04 0.46 0.23 0.01 7.0 1
Mizoram 0.17 0.02 0.07 0.14 0.38 0.20 0.02 7.3 1
Nagaland 0.19 0.03 0.16 0.07 0.26 0.28 0.02 2.3 1
Sikkim 0.20 0.01 0.17 0.03 0.10 0.48 0.02 12.1 1
Assam 0.11 0.15 0.03 0.08 0.48 0.04 0.10 7.1 2
Daman & Diu 0.15 0.06 0.05 0.18 0.46 0.03 0.07 10.0 2
Goa 0.06 0.02 0.02 0.29 0.51 0.09 0.01 22.4 2
Haryana 0.06 0.08 0.04 0.12 0.41 0.10 0.19 18.8 2
Himachal Pradesh 0.14 0.02 0.02 0.07 0.68 0.05 0.02 16.0 2
Jammu & Kashmir 0.11 0.02 0.03 0.04 0.67 0.07 0.06 9.6 2
Karnataka 0.09 0.04 0.06 0.15 0.48 0.09 0.09 15.3 2
Kerala 0.08 0.02 0.02 0.11 0.47 0.04 0.26 12.4 2
Rajasthan 0.08 0.06 0.03 0.15 0.45 0.05 0.18 13.6 2
Tamil Nadu 0.06 0.06 0.02 0.15 0.53 0.05 0.12 21.7 2
Bihar 0.13 0.20 0.03 0.11 0.14 0.03 0.36 4.7 3
India 0.07 0.10 0.04 0.15 0.34 0.06 0.24 11.4 3
Jharkhand 0.12 0.23 0.06 0.15 0.19 0.04 0.21 6.8 3
Maharashtra 0.07 0.05 0.05 0.14 0.19 0.05 0.45 12.4 3
UƩar Pradesh 0.08 0.19 0.04 0.13 0.21 0.04 0.32 7.5 3
West Bengal 0.08 0.14 0.02 0.04 0.25 0.03 0.45 6.3 3
Andhra Pradesh 0.09 0.09 0.08 0.16 0.40 0.04 0.14 17.9 4
Dadra & Nagar Haveli 0.11 0.09 0.13 0.24 0.31 0.07 0.05 17.3 4
Delhi 0.07 0.07 0.04 0.20 0.35 0.20 0.07 12.2 4
Gujarat 0.09 0.10 0.12 0.28 0.25 0.08 0.09 12.8 4
Manipur 0.10 0.10 0.11 0.14 0.44 0.09 0.02 5.5 4
Tripura 0.15 0.18 0.09 0.10 0.34 0.13 0.01 6.8 4
UƩarakhand 0.11 0.12 0.05 0.16 0.39 0.11 0.07 8.8 4
Chandigarh 0.05 0.25 0.04 0.24 0.20 0.20 0.02 12.3 5
Chhaƫsgarh 0.08 0.25 0.03 0.29 0.19 0.05 0.12 12.4 5
Madhya Pradesh 0.12 0.15 0.03 0.20 0.27 0.04 0.19 11.6 5
Odisha 0.10 0.24 0.02 0.20 0.28 0.03 0.14 9.2 5
Puducherry 0.05 0.12 0.03 0.33 0.40 0.05 0.02 18.8 5
Punjab 0.09 0.21 0.03 0.20 0.34 0.07 0.07 14.2 5
Walk-IPT-Bus
Bus
Cycle and Train
IPT and 2W
Cycle and 2W
Fig. 2. Mode shares within each state and corresponding cluster.
(Shades of green represents below average values, white represent
average, and shades of red represent above average; darker colours
represent values farther from average; titles for each cluster represent
the group of modes at above average levels in each cluster).
Table 2
Average mode share by cluster (Shaded cells represent modes with high share in the
cluster).
R. Goel Accident Analysis and Prevention 112 (2018) 105–115
109
high levels of 2W and IPT, and cluster 5 has high levels of cycle and 2W.
Cluster 2 is the only one with a high share of bus use.
The appearance of the same colour band for a given mode and
fatality rate represents a positive correlation between the two. For in-
stance, in Cluster 1, the shades of green for 2W, cycle, and train appear
with shades of green of fatality rates, thus indicating a positive corre-
lation between each of the three modes and fatality risk. In cluster 3,
the shades of green for bus and car appear with shares of green of
fatality rates, thus indicating a positive correlation of the two modes
with the fatality risk. Median fatality rate for the 5 clusters are—7.7,
14.7, 8.2, 12.3, and 13.1 fatalities per 100,000 persons for clusters 1
through 5, respectively.
It is interesting that the clustering of states using mode shares also
independently clusters states with similar fatality risk. For instance, all
the states within Clusters 1 and 3 have average or below average
fatality rates. Cluster 2 has the highest median fatality risk and consists
of 6 of the 10 states with the highest levels of fatality risk. This indicates
a correlation between mode shares and fatality risk. The Pearson cor-
relation between fatality rates and mode shares is –0.47 for walk, –0.23
for cycle, –0.26 for IPT, 0.5 for 2W, –0.19 for car, 0.35 for bus, and
–0.16 for train.
5.2. Regression analysis
Table 3 presents mean and standard deviation of the posterior dis-
tributions of regression coefficients for the four models along with their
significance based on 95% BCI. It is noteworthy that the direction of
association (sign of coefficients) of different variables with the risk
remains the same across all the four models. Thus, it is less likely that
an omitted variable is leading to a biased result as far as the direction of
association is concerned. The significance of some coefficients varies as
more variables are added. Coefficients of cycle and 2W remain con-
sistently significant across all the models, while the coefficients of walk,
cycle, 2W and diesel, are significant across all models (2, 3 and 4) ex-
cept 1.
The inclusion of diesel in model 2 has a significant effect on mag-
nitude of coefficients of walk and 2W, and to a lesser extent, on the
coefficient of IPT and cycle. The coefficient of walk increased while
those of cycle and 2W reduced. In model 3, the addition of national
highways results in the reduction of the magnitude of walk, almost
negating the increase from the addition of diesel in model 2. The effects
of proportion urban population and density in model 4 is insignificant,
however, their addition to the model results in an increase in the
magnitude of coefficient of car. The effect of car is also significant in
model 4, while it is insignificant in the other three models. Model 4 is
considered as the final model, and the rest of the discussion in the paper
will be based on results of this model.
To summarise the results of the regression model, fatality risk is
positively associated with distance travelled by 2W, car, and bus, and
negatively associated with distance travelled by walk, cycle and IPT.
The effect of bus is mixed because walking from bus also contributes to
total walking distance and the latter has a negative effect on risk. IPT
contributes to lower risk through vehicular distance (IPT has a negative
coefficient) as well as through its contribution to walking distance. In
addition, walking distance includes walking contributed by train.
Therefore, all forms of PT indirectly contribute to reduced risk.
The direction of association of modes with fatality risk is consistent
with the Pearson correlation between mode shares and risk as discussed
in Section 5.1, except in case of car. The correlation in case of car is
negative while its effect is positive in the regression model. This is likely
because high share of car is correlated with low share of 2nW. There-
fore, a negative correlation with risk is likely a reflection of low share of
2W. The regression model, on the other hand, estimates an independent
effect of car.
The effects of NH and proportion urban population are negative,
while density has a positive effect. The effect of national highways is
most counterintuitive and is expected to be in the opposite direction.
Since length of other road types has not been included, NH may be
indicating the effect of overall road network. An increase in the density
of road network may be an indicator of higher congestion. The effect of
urban population indicates higher safety resulting from slower travel
speed within urban areas, as opposed to the faster moving traffic on
rural inter-city roads.
5.3. Sensitivity analysis
I conducted sensitivity analysis of assumption of access-egress dis-
tance for all PT trips longer than 1 km (see Appendix: Table A3). In the
main analysis, distance of 1 km is assumed. For sensitivity, I assumed no
distance and 1.5 km. For no distance assumed, the effect of bus becomes
weaker and in case of 1.5 km it becomes stronger. This is expected
because when no walking distance is assumed, bus coefficient re-
presents a mix of positive effect of bus and a negative effect of walking,
thus effectively reducing its positive effect. Also, the magnitude of walk
coefficient reduces in case of no walking distance for PT. Again, the
effect of including walking distance for PT is in the direction as ex-
pected. Without this distance, the effect will be mixed with the coeffi-
cient of bus. The effect of these assumptions on IPT coefficient was
similar to that of bus.
6. Scenarios
To develop scenarios, I used average mode shares of the five clusters
(Table 2) as the five baselines representing different travel patterns
across the states in India. The scenarios have been modelled to illustrate
safety effect of mode shift among three different pairs of modes—(A)
2W and car, (B) walk and 2W, and (C) cycle and 2W. The directions of
mode shifts in the three pairs are 2W to car, walk to 2W, and cycle to
2W, respectively. The mode shift scenarios represent the gradual shift of
road users from their current mode to a new mode as they acquire a
new vehicle. For pedestrians and cyclists, the next affordable mode is
2W, and for 2W users, it is car. The shift to and from PT has not been
included. Excluding intercept and other control variables, the model I
use to develop scenarios has the form as shown in Eq. (6).
= − − −f walk cycle IPT bus W car2n 0.36 0.2 0.23 0.07 0.39 0.26 (6)
The model exponents in Eq. (6) are the mean values of coefficients
from the final model (model 4; Table 3). In the three scenarios, the
mode shift was modelled in ten steps including the baseline. In each
step, 0.5 percent points of mode share are shifted from one mode and
added to another. This ensures that the total distance remains the same.
For instance, for cluster 1, the baseline mode shares of car and 2W are
0.112 and 0.084, respectively. In the next step, the share of car is 0.117
(0.112+0.005) and that of 2W is 0.079 (0.084–0.005), and so on for
further steps.
To estimate relative risk (RR) for each incremental step, number of
fatalities ( fn) calculated for each step was divided by the fatalities
calculated for the baseline. Therefore, in effect, the RR of baseline is
one. Note that while the model was developed with total distance tra-
velled, for the scenarios, I am using mode shares, which is effectively
multiplying the original model with a constant.1 In case of Cluster 1 in
scenario C, only 4 steps are modelled as the baseline mode share of
cycle in the cluster is very low and approaches to zero by the fourth
step.
Among the three scenarios (see Fig. 3), the growth rate in RR pre-
dicted for scenario B (Walk to 2W) and scenario C (Cycle to 2W) is
much steeper than scenario A (2W to Car). RR is predicted to reach 1.3
1 For example:
= =− − −walk(walkmode share) (walk/total distance) total distance0.34 0.34 0.36 0.34, where
total distance0.34 is a constant for a given state
R. Goel Accident Analysis and Prevention 112 (2018) 105–115
110
to 1.5 in scenario B, 1.2 to 2.5 in scenario C, and 0.8 to 1.2 in scenario
A. This is expected because in scenario A, mode shift is occurring
among modes (2W and car) which are both associated with higher risk
(positive exponent). On the other hand, in scenarios A and B, mode shift
is occurring from low-risk modes with negative exponents (walk and
cycle) to a high-risk mode with positive exponent (2W).
Unlike scenarios B and C, RR does not have a monotonic trend in
scenario A. For instance, Cluster 1 experiences a decline in RR while
Cluster 5 experiences a high growth rate. The other three clusters (2, 3
and 4), on the other hand, experience a slow rate of increase and have a
slight U-shaped form. U-shape indicates a possibility of a critical point
beyond which further mode shift from 2W to car results in higher
safety. Thus, Cluster 1 consists of states which are already beyond that
critical point, while all the states in the other four clusters will reach
that point in the future. It is, therefore, worth investigating the travel
patterns in Cluster 1 in relation to other clusters.
Among all the clusters, Cluster 1 has the highest mode share of car
and the lowest share of 2W (see Table 2). The ratio of mode shares of
cars to 2W in this cluster is 3.5 while this ratio ranges from 0.3 to 0.56
in other four clusters. In Cluster 5, which experiences highest growth in
RR, mode share of 2W is the highest among all the clusters and ratio of
car to 2W share is 0.3. Clearly, both Clusters 1 and 5 stand out in terms
of their mode shares of 2W and car as well as relative ratios of the two
modes. Cluster 1 represents a system with many more cars with a small
number of 2W, and Cluster 5 represents the reverse. The same mode
shift makes Cluster 1 safer while it makes Cluster 5 less safe. Interest-
ingly, all the states in Cluster 1 are contiguous and are in the north-
eastern part of the country.
7. Discussion
In this study an ecological regression model was developed at the
state level to understand the relationship between road deaths and
commute distance travelled by different modes in India. The relation-
ship controlled for diesel consumed, length of national highways, per-
cent population urban, and population density. The regression model-
ling was carried out using hierarchical Bayesian framework thus
ensuring stable estimates with limited number of units of analysis. An
additional benefit is that coefficients are reported in terms of distribu-
tions, and therefore, a subjective understanding of the coefficients can
Table 3
Regression model.
Model 4 (Final) Model 3 Model 2 Model 1
Mean SD Mean SD Mean SD Mean SD
(Intercept) −9.178b 1.030 −8.645b 0.937 −8.113 0.951 −10.214 0.592
busln( ) 0.066 0.109 0.073 0.108 0.112 0.112 0.114 0.125
IPTln( ) −0.234b 0.113 −0.198a 0.110 −0.157 0.114 −0.109 0.125
carln( ) 0.263a 0.138 0.146 0.105 0.105 0.108 0.073 0.119
walkln( ) −0.355b 0.178 −0.281a 0.169 −0.434b 0.160 −0.275 0.156
cycleln( ) −0.200b 0.082 −0.160b 0.076 −0.152a 0.080 −0.220b 0.085
Wln(2 ) 0.390b 0.169 0.238a 0.128 0.270b 0.135 0.462b 0.128
dieselln( ) 0.264a 0.131 0.345b 0.117 0.331b 0.124
Length of NH kmln( ( )) −0.146b 0.068 −0.085b 0.041
ProportionUrban population −0.832 0.613
Density 0.039 0.108
a Significant at 90% BCI.
b Significant at 95% BCI.
Fig. 3. Relative risk for 5 clusters of Indian states resulting from incremental mode shift in three scenarios (Each step is 0.5% points shift from one mode to another).
R. Goel Accident Analysis and Prevention 112 (2018) 105–115
111
be made.
The regression model shows that walking, cycling, and IPT are as-
sociated with lower risk of road deaths in states. On the other hand, car,
2W, and bus are associated with higher risk. Thus, out of six on-road
modes of transport, half are associated with lower risk, while the other
half with higher risk. The two modes with the highest magnitude of the
coefficients are walk and 2W, however, with opposite signs. The model
thus represents the interaction of multiple modes on Indian roads and
their competing effects towards risk. The coefficients imply that any
mode shift from walking, cycling or IPT to any of the other three modes
(2W, car and bus) is likely to result in higher number of road deaths.
An interesting finding is that IPT such as auto rickshaws or tuk-tuks
are associated with lower risk of road deaths. This is in line with the
safety hypotheses of auto rickshaws reported earlier (Mohan and Roy,
2003; Pandey et al., 2015). IPT vehicles have a small engine size of
300 cm3 which limits their speed to 50–60 km/h, compared to more
than 1600 cm3 of an average car in India with much higher speed.
Further, the weight of an auto rickshaw is one fifth of a small car. With
smaller engine size and body weight, in case of a crash, auto rickshaw
results in lower injuries to pedestrians and cyclists. At the same time,
due to an enclosure, it provides safety to its occupants in case of a crash
with a car or a heavier vehicle. Low capability of speeding also results
in traffic calming of other motorised traffic. This finding has an im-
portant implication because IPT is the only motorised mode which is
associated with higher safety. At the same time, it is a form of public
transport and hence its growth is desirable from sustainable transport
perspective.
The effect of bus is mixed with bus distance associated with higher
risk of road deaths, and the walking distance (accompanied with bus
trips) associated with lower risk. While a large proportion of passenger
transportation in India is carried out through buses, the infrastructure is
insufficient for the buses to operate safely. Very often buses in India
share the road space with pedestrians and cyclists in the absence of
dedicated infrastructure for the latter two road users (Tiwari et al.,
1998). There are also bus design issues which result in higher number
of deaths. Due to absence of automatic doors, passengers are injured
from a fall during boarding or alighting. Due to high ground clearance
of bus, in case of a crash, pedestrians and cyclists are crushed under the
wheels of buses, which would be prevented with a low clearance
(Kharola et al., 2010).
The effect of bus agrees with the results reported by Bhalla et al.
(2007) who developed a risk-based injury model supported by em-
pirical estimates of case fatality ratios. The authors also reported an
increase in the number of deaths with a mode shift scenario in favour of
buses. However, the number of deaths in bus scenario was much lower
compared to those where mode shift occurred towards cars and 2W.
The regression results in the present study also indicate that the coef-
ficient of bus is positive (therefore the deaths will increase with higher
bus share) but is much lower in magnitude than that of cars and 2W.
Use of buses is desirable from the perspective of sustainable transport,
and with low levels of vehicle ownership, a large section of society
depends on buses for their daily travel. In this respect, road infra-
structure needs to be designed and bus design needs to be modified to
minimise the externalities of bus use.
Among the two private motorised modes, exponent of 2W is more
than 1.5 times higher than that of cars. This indicates that for a given
distance, 2W results in much higher risk than cars. Note that the de-
pendent variable in the model is total road deaths. Thus, the coefficient
for a vehicle type indicates risk of the vehicle occupants as well as those
struck by it. 2W are hazardous for their riders as well as those struck by
them such as pedestrians or cyclists. Cars, on the other hand, have
much higher safety for their occupants, while highly hazardous to the
colliding road users such as pedestrians, cyclists as well as 2W riders.
This is likely why the coefficient for 2W is higher than cars. This is
possibly why mode shift from 2W to car results in a U-shaped trend. As
car share increases, beyond a certain mode share mix, the effect of
lower risk of cars overcomes the higher risk of 2W, and the resulting
effect is an overall reduction in number of deaths.
U-shape trend of fatalities observed in the 2W to car scenario has
been empirically observed earlier in the road injury literature
(Söderlund and Zwi,1995; Kopits and Cropper, 2003). According to
these, road traffic injury burden in countries, often expressed as fatal-
ities per capita, rises to a certain level of per capita income and then
reduces with increasing income. Explaining this phenomenon, Bishai
et al. (2006) concluded, among other factors, that the reduction is
possibly because road users become safer as they shift from vulnerable
modes of transport to cars with much higher safety. A similar finding
was reported by Bhalla et al. (2007) using their risk-based model. In the
scenarios involving mode shift to car, the authors found a U-shaped
trend which they attributed to increasing safety of car occupants.
Scenarios of mode shift from walking and cycling to 2W present a
worrying trend of fatalities if growth in 2W ownership continues to
replace existing trips of walking and cycling. The fatalities according to
the scenarios are predicted to increase at a steep rate. If past trends are
any indicator, the steep growth of fatalities is not completely hy-
pothetical. For instance, from 1996 to 2014, a period of less than 20
years, fatality rate increased by 3 times in Punjab state and 2 times in
Chandigarh and Sikkim. Many other states also experienced high
growth rate if not as dramatic as the three (Mohan et al., 2015).
With a high share of 2W in motorised fleet, India and many other
south-Asian settings face a special challenge of road traffic injuries. For
a low-income population, 2W present an option of owning a motorised
mode with less than one-third the cost of buying a car, 3 times higher
fuel efficiency (Goel et al., 2015), and lower parking space requirement.
While 2W always have a higher risk than cars, these risks need to be
minimised by strict enforcement of motorcycle helmets. The risks can
be further reduced by the implementation of speed calming measures.
8. Strengths and limitations
The study has strengths as well as some limitations. This paper re-
ports the use of census data to develop an injury prediction model ac-
counting for exposure of all road users. For a setting like India with a
complex mix of traffic modes, this study adds a significant under-
standing of how road death burden will evolve as travel patterns change
in future. This is the first such study in India and the methods can be
applied to model injuries at city or district level.
Given that it is an ecological study with large units of analysis, the
results may be biased due to modifiable area unit problem. The number
of fatalities within a state are an aggregate of urban and rural areas. In
the rural areas, which include highways, both the type as well as the
speed of traffic is different from the urban areas. On the highways,
traffic is dominated by cars, buses and trucks, while within urban areas,
traffic consists of many more pedestrians, cyclists, and 2W users.
Though all types of road users can be seen in urban areas as well as on
highways in India as the latter often pass through village settlements or
towns. There is a possibility that distance travelled by cars, buses, and
trucks will translate into different risk to its occupants as well as other
road users in urban areas compared to rural areas.
The other limitation is the use of deaths as the measure of road
injury burden, which represent only a small fraction of total crashes. It
is also likely that while the deaths reduce in a scenario, the number of
serious injuries may still rise.
Acknowledgements
This work was undertaken under the auspices of the Centre for Diet
and Activity Research (CEDAR), a UKCRC Public Health Research
R. Goel Accident Analysis and Prevention 112 (2018) 105–115
112
Centre of Excellence funded by the British Heart Foundation, Cancer
Research UK, Economic and Social Research Council, Medical Research
Council, the National Institute for Health Research (NIHR) and the
Wellcome Trust (MR/K023187/1). Author’s contribution was also
supported by TIGTHAT, an MRC Global Challenges Project MR/
P024408/1.
Appendix A. (All data used in this manuscript is available on request from Dr Rahul Goel :rahulatiitd@gmail.com)
Table A1
Number of road deaths in each state over 2010–2012 period.
State 2010 2011 2012 Average
Arunachal Pradesh 139 126 136 134
Assam 2030 2342 2291 2221
Bihar 4693 5072 5056 4940
Chandigarh 138 136 114 129
Chhattisgarh 2888 3454 3167 3170
Dadra & Nagar Haveli 62 63 53 59
Daman & Diu 23 21 29 24
Goa 342 338 302 327
Gujarat 7384 8006 7855 7748
Haryana 5006 4681 4598 4762
Himachal Pradesh 1099 1083 1109 1097
Jammu & Kashmir 1029 1140 1426 1198
Jharkhand 2140 2053 2512 2235
Karnataka 9574 8958 9448 9327
Kerala 3950 4145 4286 4127
Madhya Pradesh 8539 8256 8506 8434
Maharashtra 14063 13680 13936 13893
Manipur 153 156 158 156
Meghalaya 184 229 213 209
Mizoram 82 81 77 80
Nagaland 44 36 56 45
Delhi 2170 2107 1866 2048
Puducherry 239 233 233 235
Punjab 2133 4897 4795 3942
Rajasthan 9163 9232 9528 9308
Sikkim 71 106 44 74
Tamil Nadu 15409 15422 16175 15669
Tripura 236 245 272 251
Uttar Pradesh 15099 14996 15109 15068
Uttarakhand 917 922 827 889
West Bengal 5470 5646 6222 5779
Odisha 4105 3797 3701 3868
Andhra Pradesh 15337 15158 14966 15154
Table A2
Average (Standard deviation) Trip Distance by mode in km for India and 33 states (Goel, 2018).
State All Modes Walk Bicycle Bus Car IPT 2W Train
India 10.1 (16.5) 2.1 (2.3) 5.4 (7.8) 21.1 (26) 15.6 (28.4) 10 (16.2) 8.2 (14.2) 51.9 (62)
Andhra Pradesh 10.5 (16.8) 2.4 (2.4) 5.1 (7.4) 23.4 (27.7) 17.0 (26.8) 10 (15.7) 8.6 (15.1) 77.7 (77.8)
Arunachal Pradesh 5.0 (12.2) 1.8 (2.4) 4.1 (6.7) 21.9 (26.3) 12.3 (27.1) 8.9 (15.4) 7.3 (14.7) 23.6 (46.4)
Assam 7.3 (13.5) 1.5 (2) 4.2 (6.4) 33.8 (33.5) 14.6 (29) 8.1 (13.5) 7.2 (13.1) 55.5 (79.7)
Bihar 9.3 (15.6) 2.4 (2.5) 6.0 (8.5) 23.9 (34.9) 16.4 (31.2) 8.9 (14.4) 9.3 (16.3) 55.1 (76.4)
Chandigarh 7.0 (11.5) 1.9 (2.1) 5.7 (7.6) 17.3 (22.8) 9.3 (15.8) 7.9 (12.3) 6.5 (9.6) 36.6 (50.4)
Chhattisgarh 7.1 (12.9) 1.5 (1.9) 5.2 (7.5) 30.1 (41.2) 17.3 (32.6) 10.5 (16.6) 8.2 (14.4) 38.1 (57)
Dadra & Nagar Haveli 4.5 (9.1) 0.9 (1.7) 3.5 (5) 12.7 (17.8) 9.5 (18) 6.7 (10.2) 6.4 (11.3) 61.4 (85.1)
Daman & Diu 5.1 (11) 1.0 (1.5) 2.7 (3.8) 32.7 (32.4) 7.3 (13.9) 6.7 (10.6) 5.6 (10) 55.1 (79.8)
Goa 10.0 (16.6) 2.0 (2.3) 5.2 (7.5) 16.2 (21.8) 13.4 (25.2) 11.6 (19.6) 10.2 (17.4) 30 (45.5)
Gujarat 7.7 (13.4) 1.9 (2.2) 4.4 (6.3) 24.2 (28.3) 14.6 (26.6) 8.5 (13.5) 6.7 (11.4) 41.3 (57.8)
Haryana 12.9 (19.2) 2 (2.3) 5.4 (7.6) 38.9 (36.7) 17.7 (31.3) 10.2 (15.7) 8.4 (14.6) 53.3 (69.7)
Himachal Pradesh 9.1 (17.8) 2.1 (2.2) 5.4 (7.6) 19.2 (24.5) 13.1 (26) 13.5 (22.4) 9.1 (15.8) 40.9 (64.4)
Jammu & Kashmir 12.5 (23.2) 2.6 (2.5) 6.3 (8.7) 22.4 (26.9) 16.3 (28.7) 12.8 (20.2) 8.8 (14.4) 66.2 (88.6)
Jharkhand 7.9 (13.8) 2.2 (2.2) 6.3 (8.6) 29.5 (41.9) 13.8 (26.1) 9.4 (14.3) 7.5 (12.8) 55.7 (76.3)
Karnataka 10.6 (16.9) 2.2 (2.3) 5.7 (8.1) 18.4 (23.8) 15.7 (27.9) 12.6 (20.3) 8.6 (14.8) 64.8 (70.1)
Kerala 8.7 (15.8) 1.7 (2) 4.3 (6.1) 12.4 (17.6) 11.1 (20.8) 6.4 (11.4) 7.7 (12.4) 81.4 (99.4)
Madhya Pradesh 7.5 (13.4) 2.2 (2.3) 5.5 (7.8) 26.4 (29.8) 15.3 (29.5) 9 (14.6) 7.5 (13.4) 76.6 (77)
Maharashtra 10.2 (18.5) 2.2 (2.3) 5.0 (7.2) 17.0 (22.6) 14.5 (22.1) 10.3 (17.2) 8.8 (15.2) 38.7 (55.2)
Manipur 8.3 (16.2) 2.1 (2.6) 5.3 (7.7) 25.8 (35.2) 13.5 (26.3) 8 (13.2) 6.8 (11.5) 23.5 (43.7)
Meghalaya 8.5 (15) 1.8 (2.2) 4.7 (7.1) 36.0 (34.2) 14.8 (28.8) 6.9 (11.9) 9.6 (17.5) 26.5 (50.9)
(continued on next page)
R. Goel Accident Analysis and Prevention 112 (2018) 105–115
113
References
Bhalla, K., Ezzati, M., Mahal, A., Salomon, J., Reich, M., 2007. A risk‐based method for
modeling traffic fatalities. Risk Anal. 27 (1), 125–136.
Bishai, D., Quresh, A., James, P., Ghaffar, A., 2006. National road casualties and eco-
nomic development. Health Econ. 15 (1), 65–81.
Census-India, 2012. Population Enumeration Data (Final Population). Accessed online.
http://www.censusindia.gov.in/2011census/population_enumeration.html.
Census-India, 2017a. B-28 'Other Workers' By Distance From Residence To Place Of Work
And Mode Of Travel To Place Of Work - 2011. Accessed online. http://www.
censusindia.gov.in/2011census/B-series/B_28.html.
Census-India, 2017b. HH-12 Number Of Households Availing Banking Services And
Number Of Households Having Each Of The Specified Assets. Accessed online.
http://www.censusindia.gov.in/2011census/Hlo-series/HH12.html.
Census-India, 2017c. B-1 Main Workers, Marginal Workers, Non-Workers and Those
Marginal Workers, Non-workers Seeking/Available for Work Classified by Age and
Sex. Accessed online. http://www.censusindia.gov.in/2011census/B-series/B-
Series-01.html.
Census-India, 2017d. Census of India 2011—Metadata. Accessed online. http://www.
censusindia.gov.in/2011census/HLO/Metadata_Census_2011.pdf.
DiMaggio, C., 2015. Small-area spatiotemporal analysis of pedestrian and bicyclist in-
juries in New York City. Epidemiology 26 (2), 247–254.
Elvik, R., 2016. Safety-in-numbers: estimates based on a sample of pedestrian crossings in
Norway. Accid. Anal. Prev. 91, 175–182.
Elvik, R., Bjørnskau, T., 2017. Safety-in-numbers: a systematic review and meta-analysis
of evidence. Saf. Sci. 92, 274–282.
Goel, R., 2017. Public Health Burden of Transport in Delhi. PhD Thesis. Transportation
Research and Injury Prevention Programme, Indian Institute of Technology, Delhi.
Goel, R., 2018. Distance-decay Functions of Travel to Work Trips in India. Data in Brief.
Submitted. .
Goel, R., Guttikunda, S.K., 2015. Evolution of on-road vehicle exhaust emissions in Delhi.
Atmos. Environ. 105, 78–90.
Goel, R., Tiwari, G., 2016. Access–egress and other travel characteristics of metro users in
Delhi and its satellite cities. IATSS Res. 39 (2), 164–172.
Goel, R., Guttikunda, S.K., Mohan, D., Tiwari, G., 2015. Benchmarking vehicle and pas-
senger travel characteristics in Delhi for on-road emissions analysis. Travel Behav.
Soc. 2 (2), 88–101.
Goel, R., Mohan, D., Guttikunda, S.K., Tiwari, G., 2016. Assessment of motor vehicle use
characteristics in three Indian cities. Transp. Res. Part D: Transp. Environ. 44,
254–265.
Goel, R., Jain, P., Tiwari, G., 2018. Correlates of fatality risk of vulnerable road users in
Delhi. Accid. Anal. Prev. 111, 86–93.
Goodman, A., 2013. Walking, cycling and driving to work in the English and Welsh 2011
census: trends, socio-economic patterning and relevance to travel behaviour in gen-
eral. PloS One 8 (8), e71790.
Hox, J.J., van de Schoot, R., Matthijsse, S., 2012. How few countries will do? Comparative
survey analysis from a Bayesian perspective. July. Surv. Res. Methods 6 (2), 87–93.
Hsiao, M., Malhotra, A., Thakur, J.S., Sheth, J.K., Nathens, A.B., Dhingra, N.,
Collaborators, f. t. M. D. S, 2013. Road traffic injury mortality and its mechanisms in
India: nationally representative mortality survey of 1.1 million homes. BMJ Open 3
(8). http://dx.doi.org/10.1136/bmjopen-2013-002621.
Kharola, P.S., Tiwari, G., Mohan, D., 2010. Traffic safety and city public transport system:
case study of Bengaluru, India. J. Public Transp. 13 (4), 4.
Kopits, E., Cropper, M., 2003. Traffic fatalities and economic growth. World Bank Policy
Res. Work. Pap., 3035.
Kumar, M., Singh, S., Ghate, A.T., Pal, S., Wilson, S.A., 2016. Informal public transport
modes in India: a case study of five city regions. IATSS Res. 39 (2), 102–109.
Mohan, D., Roy, D., 2003. Operating on three wheels: auto-rickshaw drivers of Delhi.
Econ. Polit. Wkly. 177–180.
Mohan, D., Tiwari, G., Bhalla, K., 2015. Road Safety in India- Status Report. New Delhi:
Transportation Research and Injury Prevention Programme. Indian Institute of
Technology, Delhi.
Mohan, D., Tiwari, G., Mukherjee, S., 2016. Urban traffic safety assessment: a case study
of six Indian cities. IATSS Res. 39 (2), 95–101.
MoPNG, 2012. Indian Petroleum and Natural Gas Statistics (2011–2012). Economic
Division, Ministry of Petroleum and Natural Gas, Government of India, New Delhi.
MoRTH, 2011. Review of the Performance of State Road Transport Undertakings
(SRTUs)—Passenger Services for April 2010- March 2011. Transport Research Wing,
Ministry of Road Transport and Highways, Government of India, New Delhi, India.
MoRTH, 2013. Basic Road Statistics of India 2011-12. Transport Research Wing, Ministry
of Road Transport and Highways, Government of India, New Delhi, India.
Naqvi, H.M., Tiwari, G., 2017. Factors contributing to motorcycle fatal crashes on na-
tional highways in India. Transp. Res. Procedia 25, 2089–2102.
NCRB, 2011. Accidental Deaths & Suicides in India 2010. Accessed online. National
Crime Records Bureau, New Delhi. http://ncrb.nic.in/.
NCRB, 2012. Accidental Deaths & Suicides in India 2011. Accessed online. National
Crime Records Bureau, New Delhi. http://ncrb.nic.in/.
Table A3
Sensitivity analysis of access distance of PT.
Final model Model assuming no distance for access-egress of PT Model assuming access-egress distance of 1.5 km for PT
Mean SD Mean SD Mean Mean
(Intercept) −9.178b 1.030 −9.017b 1.065 −9.251 1.020
busln( ) 0.066 0.109 0.034 0.102 0.077 0.112
IPTln( ) −0.234b 0.113 −0.220a 0.116 −0.240b 0.113
carln( ) 0.263a 0.138 0.248a 0.137 0.267a 0.139
walkln( ) −0.355b 0.178 -0.336a 0.171 −0.356b 0.180
cycleln( ) −0.200b 0.082 −0.187b 0.084 −0.205b 0.081
Wln(2 ) 0.390b 0.169 0.373b 0.170 0.396b 0.169
dieselln( ) 0.264a 0.131 0.269b 0.132 0.262b 0.131
Length of NHs kmln( ( )) −0.146b 0.068 −0.139b 0.069 −0.149b 0.068
ProportionUrban population −0.832 0.613 −0.855 0.619 −0.818 0.612
Density 0.039 0.108 0.044 0.108 0.038 0.108
a Significant at 90% BCI.
b Significant at 95% BCI.
Table A2 (continued)
State All Modes Walk Bicycle Bus Car IPT 2W Train
Mizoram 4.3 (9.7) 0.9 (1.8) 3.4 (5.5) 8.8 (14.8) 10.3 (23.8) 4.3 (8.5) 4.5 (8.8) 24.8 (49.9)
Nagaland 4.6 (10.2) 1.2 (2) 3.2 (4.9) 13.4 (19.4) 10.4 (22.4) 6.7 (12.1) 5.7 (10.7) 21.6 (44.7)
Delhi 8.8 (14.2) 1.6 (1.9) 5.9 (8) 12.0 (15.7) 14.0 (18.6) 13.5 (21.6) 10.9 (17.4) 19 (23.8)
Odisha 9.5 (15.8) 2.2 (2.3) 6.2 (8.9) 33.6 (46) 19.7 (36.3) 12 (19.3) 9.9 (17.4) 60 (82.5)
Puducherry 8.8 (14.6) 1.9 (2.2) 4.8 (6.9) 18.1 (23.6) 14.9 (28.2) 9.1 (15.9) 7.1 (12) 38.1 (59.5)
Punjab 8.8 (14.9) 2.3 (2.5) 5.8 (8.2) 30.8 (32.5) 13.3 (24.5) 9.1 (14.6) 7 (12.2) 47.4 (69.7)
Rajasthan 10.4 (16.8) 1.9 (2) 5.1 (7.1) 32.8 (33.4) 15.8 (28.7) 9.4 (14.7) 7.9 (13.4) 75.2 (90.8)
Sikkim 5.6 (12.3) 1.7 (2.1) 7.0 (10.1) 14.6 (20.4) 12.6 (24.2) 9.9 (16.9) 9.7 (16.4) 36.6 (61.6)
Tamil Nadu 11.7 (18) 1.9 (2.2) 4.5 (6.6) 20.4 (25.5) 16.5 (29.4) 14.2 (23) 8 (13.6) 51.4 (61.6)
Tripura 6.1 (11.7) 1.7 (2) 3.9 (5.6) 24.3 (33.3) 16.9 (24.5) 7.2 (11) 7 (12.1) 18.9 (34.4)
Uttar Pradesh 10.6 (17) 2.3 (2.5) 6.6 (9.3) 36.1 (34.8) 16.3 (25.9) 10.9 (17.1) 9.6 (16.4) 85.2 (81.8)
Uttarakhand 8.3 (14.4) 1.9 (2.2) 5.5 (7.8) 29.5 (31.6) 16.6 (30.5) 9.2 (14.1) 7.3 (12.8) 38.5 (60.5)
West Bengal 12.1 (18.5) 2 (2.4) 5.0 (7.5) 21.4 (26.2) 11.7 (21.8) 10.5 (17.6) 8.5 (15) 41.2 (54.7)
R. Goel Accident Analysis and Prevention 112 (2018) 105–115
114
NCRB, 2013. Accidental Deaths & Suicides in India 2012. Accessed online. National
Crime Records Bureau, New Delhi. http://ncrb.nic.in/.
Nielsen, 2013. All India Study on Sectoral Demand of Diesel and Petrol. Commissioned by
Petroleum Planning and Analysis Cell. Ministry of Petroleum and Natural Gas,
Nielsen (India) Pvt. Ltd, Delhi, India.
NRSC, 2016. Bhuvan Thematic Services. Accessed online. Bhuvan, Indian Geo-Platform
of ISRO, National Remote Sensing Centre, Bangalore, India. http://bhuvan.nrsc.gov.
in/gis/thematic/index.php.
Pandey, G., Mohan, D., Rao, K.R., 2015. Why do three-wheelers carrying schoolchildren
suffer very low fatal crashes? IATSS Res. 38 (2), 130–134.
Rue, H., Martino, S., Lindgren, F., 2009. INLA: Functions Which Allow to Perform a Full
Bayesian Analysis of Structured (Geo-) Additive Models Using Integrated Nested
Laplace Approximation. R Package Version 0.0 ed.INLA: Functions Which Allow to
Perform a Full Bayesian Analysis of Structured (Geo-) Additive Models Using
Integrated Nested Laplace Approximation. R Package Version 0.0 ed.
Schepers, J.P., Heinen, E., 2013. How does a modal shift from short car trips to cycling
affect road safety? Accid. Anal. Prev. 50, 1118–1127.
Söderlund, N., Zwi, A.B., 1995. Traffic-related mortality in industrialized and less de-
veloped countries. Bull. World Health Organ. 73 (2), 175.
Statista, 2017. Percentage of Households Owning a Car in Selected Countries in 2014, by
Country. Accessed online. https://www.statista.com/statistics/516280/share-of-
households-that-own-a-passenger-vehicle-by-country/.
Tiwari, G., Mohan, D., Fazio, J., 1998. Conflict analysis for prediction of fatal crash lo-
cations in mixed traffic streams. Accid. Anal. Prev. 30 (2), 207–215.
R. Goel Accident Analysis and Prevention 112 (2018) 105–115
115