Content uploaded by Stefan Flügel
Author content
All content in this area was uploaded by Stefan Flügel on Mar 29, 2019
Content may be subject to copyright.
Working Paper 51396 Oslo, 8.1.2019
4530 VERDSETT, 4329 SIS-Modeller
Stefan Flügel, Askill H. Halse, Nina Hulleberg, Guri Natalie Jordbakke
Estimating the effect of vehicle
automation on car drivers’ and car
passengers’ valuation of travel time
savings
Paper presented at the 41th Annual Meeting of the Norwegian Association of Economists in Tromsø
Abstract
The Value of Travel Time Savings (VTTS) is a central parameter in transport economics
and travel demand modelling. It is expected that vehicle automation will decrease travellers’
VTTS on roads due to the possibility of performing more pleasant and useful activities
while driving (e.g. reading and sleeping).
The VTTS of autonomous vehicles (AV) is a relatively young research topic. For the time
being, it is inherently limited to stated preference studies. In contrast to the few published
international papers on the topic, our method relies on route choice experiments (RCE).
This allows us to model VTTS directly and without the danger of confounding effects
between the travel time parameter and the alternative specific constants.
We use recent data from the Norwegian Valuation Study where car drivers and car
passengers went through RCE not only in the current scenario of conventional cars but
also in different scenarios of vehicle automation. With a mixed logit model on pooled data,
we estimate the relative reduction in VTTS to be around 21% for car drivers (taking the
average effect over 4 different levels of AV). The reduction is statistically significant as
demonstrate with Likelihood Ratio tests. We also find that car users have lower VTTS in
private full automated cars compared to partial automated vehicles. We also get an
indication for that the VTTS in shared AV is somewhat higher than in private AV.
Keywords: autonomous vehicles, value of time, value of travel time saving, route choice experiments, degrees
of automation
1 Introduction
The Value of Travel Time Savings (VTTS) is a central parameter in transport economics
and travel demand modelling. It is defined as traveller’s willingness-to-pay for marginal
reductions in travel time and is typically expressed in NOK/hour. In Norway, the current
handbook on cost-benefit analyses for road projects recommends the following values for
the VTTS in cars (Table 1).
Table 1: Recommended unit values for VTTS in cars in Norway (2016-values)
Trip distance
0-70km
Over 70 km
Commuting
100 NOK/h
217 NOK/h
Business
449 NOK/h
449 NOK/h
Other trips (leisure)
85 NOK/h
169 NOK/h
*Values based on handbook of the Norwegian Public Road administration (SVV)
The VTTs is thus segmented by trip purpose and trip distance but not by car type. One
currently also assumes the same VTTS for car drivers and for car passengers. This is
despite some empirical evidences suggesting that the VTTS for car passengers is around
75% of the VTTS of car drivers (Ho et al. 2015).
It is expected that vehicle automation will increase the comfort level (in a wide sense) of
travelling due to the possibility of performing more pleasant and useful activities while
driving (e.g. reading and sleeping) and thus reduce VTTS. A possible (large) reduction in
VTTS can have strong effects on travel demand and travel mode choice as well as on the
social profitability of infrastructure projects. There is therefore an increasing interested in
estimated the expected VTTS in autonomous vehicles (AV).
The VTTS of AV is a relatively young research topic. For now, it is inherently limited to
stated preference studies as high/full automation is not yet experienced by respondents.
This makes the use of revealed preference an unavailable option.
There is no prior evidence on the effects of vehicle automation on VTTS in Norway. Using
a simple questionnaire, Flügel et al (2017) find that (only) 34.1% of current car drivers
(44.8% of all car passengers) would have preferred to take an AV over a conventional car
on their last car trip. How this result relates to the VTTS in AV is however unclear as the
reasons of not choosing AV was not further investigated in that study.
1
While there are many papers that discuss possible impacts on vehicle automation in general
(e.g. Bahamonde-Birke et al 2016)) and on VTTS (most recently Fosgerau 2018), only a few
studies have empirically estimated the effect on VTTS.
1
As the share choosing AV was significantly lower for elder people, it was suggested that conservatism was
likely to have been a contributing factor for the obtained results.
Using SP-data from Germany, Steck et al (2018) find a 31% reduction in commuters’
VTTS for private – unshared – AV, and a 10% reduction for shared AV. A reduction in
VTTS is found for all income groups (low, middle, high) and for two different model
specifications. It is, however, not clear if these reductions are statistically significant.
Using SP-data from the Netherlands, Looff et al (2018) finds that an VTTS in an
autonomous vehicle with “office interior” has a 25.2% reduction in VTTS compared to
conventional cars
2
. Interestingly and somewhat surprisingly, the autonomous vehicle
specification with “leisure interior” gave a higher VTTS than in conventional cars. As in
Steck et al (2018), no significant tests are presented.
In another SP-study from the Netherlands, Yap et al (2016) finds that VTTS in
autonomous vehicles to be considerable higher than with manual driving. The context for
AV was access mode to public transport, which is clearly different from AV as a main
mode. Still is a surprising result, as one would generally expect a reduction of VTTS in AV
scenarios (see discussion in section 3).
From a methodological point of view, it is interesting to note that all – to us known –
papers on the VTTS in autonomous cars use mode choice experiments to elicit VTTS. For
instance, in De Looff et al (2018), the choice is between a conventional car and two types
of AV (one with office interior and one with leisure interior). In Steck et al (2018), the
choice is between walk, bike, public transport, autonomous vehicles and shared
autonomous vehicles. A general challenge in mode choice analysis is that the alternative
specific constant (ASC) in the utility functions will partially absorb the comfort difference
of transport mode. This comes at the danger of confound effects between the ASC and the
marginal utility of travel time. Another issue may be, that mode choice experiments
including a not-yet-experienced travel alternatives can easily be perceived as a general
voting for or against this new travel option, rather than a choice based on the presented
attribute values.
In contrast to the few published papers, our method relies on route choice experiments
(RCE) where travellers in a first sequence of RCE choose between two routes in
conventional cars and then – in a second round of RCE – between two routes in
autonomous cars. In these (unlabelled) RCE, there is no problem with alternative specific
constants that could potentially influence VTTS-estimates.
Another advantage of our study is the large data set we can use for parameter inference and
hypothesis testing. Our estimation models are based on around 30 000 SP-route choices,
collected in the 2018-Norwegian valuation study (see more information in section 5). The
large data collection and the survey design allows us to differentiate between current car
drivers and current car passengers as well as 4 different degrees of vehicle automation.
Both aspects add to the literature of VTTS in autonomous vehicles.
2
Based on the mixed logit results, 8.37 Euro/hour in conventional car and 6.26 Euro /hour in autonomous
cars.
2 Theory
Time allocation modes (Becker 1965, DeSerpa 1971, Jara-Diaz and Guevara 2003) are the
microeconomic foundation of VTTS. The most common decomposition of individual q’s
VTTS going back to Oort (1969) is here restated with the notation used in Flügel (2014):
(1)
where is the nominal wage rate of individual q,
the marginal utility of work time,
the marginal utility of travel time and the marginal utility of income.
represents the opportunity cost of travelling (OCT) and may be assumed to be independent
of used transport mode (
) but it is likely to differ across user groups g
(
) due to differences in average income. The latter suggests self-selection of
high-income persons to fast transport modes. For instance, air travellers are likely to have a
higher VTTS than long-distance bus users as the former (latter) groups has revealed high
(low) opportunity cost of time by their choice of the fast (slow) transport mode air (bus).
is likely to depend on the specific travel mode (
due to difference in
comfort and the possibility to conduct useful activities.
Following Flügel (2014), we express the VTTS in travel mode k of group , i.e. the
current users of travel mode k, as:
(2)
.
where
are representative values for the user group .
Further, we define the mode effects (ME) between travel mode k and travel mode l (here
for user group ) as
(3) .
Mode effects can be calculated for any group (subgroup) of persons if one has information
on the VTTS in (at least) two travel modes. Note that in general.
Inserting (2) in (3) and simplifying, we can write the mode effects as:
(4)
.
While will have an effect on the absolute size of ME (but not on its sign), equation
(4) shows that the sign of the mode effect depends on the (direct) disutility obtained from
time spent in the two travel modes. A positive mode effect implies that travel
mode k in - comparison to travel mode l - provides a less useful/pleasant travel time. As
discussed in the next section, the usefulness of travel time is likely to depend on quality of
secondary activities that can be performed while travelling.
3 Degrees of vehicle autonomation and
hypotheses on VTTS
Various types and definition of degrees (levels) of vehicle automation have been proposed
(e.g. Kyriakidis et al (2015)). We depart from the definitions given in Gasser and Westhoff
(2012) which are cited below:
Partially automated driving = The automated driving system takes over both speed
and steering control on all roads. However, the system cannot handle all possible
situations. Therefore, the driver shall permanently monitor the road and be
prepared to take over control at any time.
Highly automated driving = The automated driving system takes over both speed
and steering control on all roads. The driver is not required to permanently monitor
the road. If automation cannot handle a situation it provides a take-over request,
and the driver must take-over control with a time buffer of 7 seconds.
Fully automated driving = The system takes over speed and steering control
completely and permanently, on all roads and in all situations. The driver sets a
destination via a touchscreen. The driver cannot drive manually, because the
vehicle does not have a steering wheel.
In this paper, we are less concerned with the technical issues behind the different degrees
of automation – those may or may not be understand by the respondents. For the effect of
the VTTS, we expect the extent to which (secondary) activities can be performed while
sitting in the car to be most prominent.
Table 2 gives the author’s assessment on the possibility to perform different activities in
different car modes. While listing to the radio is possible in all modes, watching movies
(and sleeping) requires high (full) automation. For current car passengers, the spectrum of
possible activities is independent on the car technology.
It is important to recall that not all activities are feasible, even with full automation (playing
tennis is mentioned as an example in table 2). This implies that VTTS is unlikely to become
zero as there will be positive cost of re-scheduling activities. In the (hypothetical) case of
no re-scheduling costs, it may be that VTTS is zero within reasonable travel times
variations. E.g. a person who has no preferences when to sleep (just how much), may have
no willingness-to-pay to reduce travel time in full automated cars, as this person could
substitute the won time with other useful activities (e.g. playing tennis) at times he/she
would otherwise sleep. Of course, the VTTS would – even in case of no re-scheduling cost
– be positive if the comfort level of the car would decrease the quality of the sleep
(compared to sleeping in a bed). The same applies for the productivity to work in the car
(compared to the office).
Table 2: Possibility to perform selected activities in different car modes
Car “driver”
Car
passenger
Activities
Convent-
ional car
Partial
automation
High
automation
Full
automation
Listening to
radio
Yes
Yes
Yes
Yes
Yes
Looking at the
landscape
Mainly no
Mainly yes
Yes
Yes
Yes
Reading/writing
messengers
No
Mainly no
Yes
Yes
Yes
Watching
movies
No
No
Yes
Yes
Yes
Sleeping
No
No
No
Yes
Yes
Playing tennis
No
No
No
No
No
It is uncertain how autonomous cars enter the future transport market. A likely scenario is
that they mainly enter in form of taxi-services (mobility-as-a-service). In this case, a person
– without paying additional charges - may have a share the ride with other persons going
in the same direction (so-called ride-sharing). It is possible that there is a difference
between the VTTS in private autonomous cars and shared autonomous cars (see discussion
below).
Safety concerns may also impact VTTS but the direction is unclear. In our questionnaire,
we asked the respondents to assume that AV are as safe as conventional cars (see details in
next section).
Considering Table 2, we expect most mode-effects between car modes to differ.
Figure 1 summaries the modes effects that we can identify with our experimental design.
3
Figure 1 also shows the notation and abbreviation that are used in the reminder of the
paper. Note that ( is the subgroup of car drivers (passengers) that have car
passenger (car drivers) as an alternative. As there may be self-selection to alternative
modes, is likely to differ somewhat in average characteristics from the
3
Technically, we estimate not “absolute” mode effects (as in equation 3) but “relative” mode effects
( , see section 6). In addition to the mode effects shown in Figure 1, we can
also derive “indirect” mode effects between two types of vehicle automation (within one each user group, car
drivers and car passengers) as a combination of the mode effects (e.g.
) = ). To avoid further confusion
however, we use the “classical” definition of the mode effects as stated in equation 3 for the formation of hypotheses.
overall car driver (car passenger) and may therefore have different preferences for travel
time savings.
Figure 1: (Main) mode effects and abbreviation used.
In the reminder of the chapter we lay out the hypotheses that we want to test in this paper.
Hypothesis 1: For conventional cars, car driving (CCD) provides less useful time
compared to being car passenger (CCP). With this hypothesis, we expect
and . The rationale behind the hypothesis is that
one can perform more useful activities as a car passengers (CCP) compared to being a car
driver in conventional cars (CCD). This is expected to be true for both relevant user
groups, i.e. car divers with car passenger as an alternative (CD_CP) and car passengers with
car driving as an alternative (CP_CD).
Hypothesis 1 is formally tested by the following two null-hypotheses (both which we
expect to be rejected).
(5)
(6)
Hypothesis 2: Current car drivers have a lower VTTS in autonomous cars. Similar as
for hypothesis 1, the intuition is that the possibility of performing more useful activities
yields lower VTTS in AVs than in conventional car driving (
for l=PAP, HAP, FAP, FAR). This is tested simultaneously by the following Null-
hypothesis (which we expect to be rejected).
(7)
Hypothesis 3: Car passengers have similar VTTS in conventional cars and fully
autonomous cars. We test the same hypothesis as for car driver, formalized by the
following null-hypotheses:
(8)
As vehicle automation does not increase the possibility to perform activities for current car
passengers (Table 2), we expect car passengers have statistically similar VTTS in all types of
cars. An exception might be that car passenger dislike ride sharing (sharing space with
potentially unknown persons) compared to private cars (where they typically share space
with people they know); see hypothesis 5.
Hypothesis 4: Higher degree of private vehicle automation yields lower VTTS for
current car drivers. As increased vehicle automation does increase the possibility to
perform more activities for current car drivers (see Table 2), we expect car drivers to have
lower VTTS in full automation compared to high (and partial) automation and lower VTTS
for high automation compared to partial automation (i.e.
).
Hence, we expect that the following null-hypothesis are statistically rejected:
(9)
(10)
(11)
Hypothesis 5: Car drivers and car passengers have lower VTTS in full automated
cars when it’s a private ride compared to ride-sharing. Our expectation is based on
results in Steck et al (2018) as well on a general intuition that many (most) persons would
prefer not to share the autonomous car with strangers and that the usefulness of time
spend in autonomous cars may be deteriorated in the presents of strangers as a
consequence of a reduction in privacy. Hence, we expect that the following Null-
hypotheses (
for car drivers and
for car passengers) are statistically rejected:
(12)
(13)
We use likelihood ratio test to test these hypotheses by comparing the Final-Log-likelihood
test statistics of a flexible model specification against different constrained model
specification (see results and discussions in section 8).
4 Experimental design
To identify the mode effects on VTTS described above, we use data from choice
experiments in which respondents make repeated hypothetical choices between alternatives
with different travel time and travel cost. The travel mode is the same in both alternatives
(‘route choice’). We combine data from three route choice experiments (RCE):
1. Choice between two alternatives using the current travel mode (car driver or car
passenger) – CE1
2. Choice between two alternatives using an alternative travel mode (car driver or car
passenger) – CE2d
3. Choice between two alternatives using an autonomous car as travel mode – CE2e
In each RCE, the respondent faces eight consecutive choice cards. An example of a choice
card is shown in Figure 2.
Figure 2. Example of choice card in the choice experiments
In the RCE involving autonomous cars, the introduction text prior to the choice cards
differs randomly among respondents.
The different versions are as shown in Box 1.
Box 1: Introduction text for the four types of vehicle automation prior to choice experiment sequence
(translated from Norwegian)
Imagine a future where cars will accelerate, slow down and control themselves.
[
For partial automation:
You do not have to drive the car i.e. you can take your hands of
steering wheel and move feet from the pedal. But you have to monitor your cars driving
and must be prepared to intervene at any time.
There is nobody in the car other than yourself and the traveling companion you had on the
trip you have described earlier.]
[
For high degree of automation
: You do not have to drive the car i.e. you can take your
hands of steering wheel and move feet from the pedal. This allows you to spend time in the
car reading or using mobile phone/pc. In rare cases, the care will notify you and you will
have to intervene (within 7 seconds) and drive yourself.
There is nobody in the car other than yourself and the traveling companion you had on the
trip you have described earlier.]
[
For full automation, private
: The car does not have a steering wheel or pedal and you
communicate with your car through a touchscreen. The car will take you to your
destination. While the car is driving, you can spend time reading, using mobile phone/pc or
sleep for example.
There is nobody in the car other than yourself and the traveling companion you had on the
trip you have described earlier.]
[
For full automation, shared
: The car does not have a steering wheel or pedal and you
communicate with your car through a touchscreen. The car will take you to your
destination. While the car is driving, you can spend time reading, using mobile phone/pc or
sleep for example.
Known or unknown people can possibly be picked up or be set off, along the way. You
and any travel companion, is guaranteed seating but it may be other people next to you.]
Assume that the safety with this car is as safe as cars today.
Imagine that you are going to undertake the journey you’ve described over. Now, the car
you use will run itself, as described above.
Two trips will now be presented to you, and those trips will differ in travel time and cost.
Assume that you have to pay the differences in cost between the two alternatives.
Your task is to choose the alternative you prefer.
Unfortunately, the introduction text is somewhat ambiguous for car passengers as it is not
clearly stated if the car passengers remain in the passenger’s seat or if they should think of
seating in the drivers’ seat where they would need to monitor the AV in partial (and high)
automation.
While the routing to CEe and to the levels of autonomations within CEe is random (and
thereby exogenous), the travel mode in CE1 is based on the actual mode choice in the
respondent’s reference trip, i.e. a trip that respondent recently made
4
.
To make the SP-choice situations in CE1 as realistic as possible, the values of attributes
travel time and cost presented in the choice cards are based on reference values
representing the characteristics of the reference trip. Respondents are asked to assume that
reference values for the alternative mode (in CE2d and CE2e) are the same as in the
reported reference trip.
The statistical design of the choice experiments is similar to the one used in the 2009 value
of time study in Norway (Ramjerdi et al 2010). Each respondent gets 2 willingness-to-pay
(WTP), 2 willingness-to-accept (WTA), 2 equivalent loss (EL) and 2 equivalent gains (EG)
choices. The changes in the time attribute (delta T) is randomly chosen between 10-30% in
the design and the difference in the cost attribute is such that it corresponds to the implicit
trade-off which is by the design randomly drawn from predefined interval ranging between
10-750 NOK/h for short trips, 10-1000 NOK/h for long distance trips and 50-1250
NOK/h for business trips (both short and long distance). The actual trade-offs (“bids”) in
the choice cards can deviate from these design-bids due to rounding and restrictions on the
attribute values.
All those who have a trip by car as their reference trip first participate in the RCE involving
their current mode of travel (CE1). They are then randomly allocated to different choice
experiments as the second one, including the two choice experiments involving an
alternative travel mode (CE2d and CE2e). Those who are allocated to CE2d are asked to
report which alternative modes were available on their reference trip, and one out of these
is selected randomly as the alternative mode. To increase combinations between car and
public transport modes (being important for other analysis in the valuation study), car
driving (being a car passenger) as an alternative to being a car passenger (car driving) was
given low priority. For our analysis, it is relevant to know that the alternative mode in
CE2d is partially endogenous as the respondents that get both RCE with car driving and as
car passenger are those respondents stating that they had no other alternative transport
mode available (which mainly applies in rural areas).
Figure 3 illustrates the structure of the questionnaire (omitting common questions about
socioeconomic variables at the end of each questionnaire). Those sampled that are
underscored are those used in this paper.
4
The reference trip is selected in one out of the two following ways: a) In the beginning of the questionnaire,
respondents are asked whether they have made a trip of at least 70 kilometers during the last two weeks. A
random share of those who answered yes are then asked to give more information about the last such trip,
which will be the reference trip. (b) The rest of the respondents are asked to list all their trips made on the
day before (‘trip diary’). One of these trips is then selected randomly as the reference trip, and the respondent
is asked to give more information about this trip.
Figure 3. Structure of the CE-related part of the questionnaire.
5 Data collection and processing
We use data collected in October and November 2018 as part of the Norwegian valuation
study on personal travel. Respondents were recruited from (1) from an email database
owned by the Norwegian postal service and (2) a commercial internet panel and invited to
fill out an on-line questionnaire.
5
A pilot, using (1), was conducted in August 2018, but this
data is not included in our analysis as it involved some programming mistakes in the
experimental design.
The response rate in sample (1) was 2.7% (completed questionnaires) as share of the total
number of email invitations sent. Since not everyone received the email or were aware that
they had received it, the effective response rate will be somewhat higher. In sample (2),
13.6 % of those invited completed the questionnaire. Vehicle automation was not
mentioned in the invitation emails, such that potential sample selection bias (given the
relative low response rate) is not likely related to preference for or against vehicle
automation.
For the analysis of this paper we have excluded the following respondents and
observations
Respondents with questionable choice behavior (4.61% of the original
sample)
In all 8 choices picked to the left-hand-side or the ride-hand-side
alternative
Respondents took less than 25 seconds to go through one of the
CE (i.e. less than 3 sec. for each choice card on average)
Respondents that finished to whole questionnaire in less than 6
minutes
5
The Norwegian valuation study also used field recruitment/interviews but not for car users. Some few
respondents recruited in public transport or on a bicycle trips did report a reference trip for car, but those are
omitted in the analysis in this paper.
Questions about reference trip
CE1 Car driver
CE2d
CE2d car
passenger
CE2e
CE2d other
modes
CE2 other
CE1 Car passenger
CE2d
CE2d car
driver
CE2e
CE2d other
modes
CE2 other
CE1 Other modes
CE2 other
modes
Respondents with missing or zero entries in central background variables:
distance, base cost, age and travel distance (0.1% of the remaining sample)
Very long reference trips (1.28% of the remaining sample)
Travel time over 24 hours
Travel distance over 1500 km
Choices where the experimental design did not work as intended (2.78% of
remaining sample)
Identical time attributes (delta T=0)
Bid = 0 NOK/h
Bid > 1500 NOK/h
Car drivers as a reference mode for respondents under 18 years (0.26% of
the remaining sample)
After these exclusions, we have 30556 SP-choices from 3111
respondents; the majority (82.3%) of the choices are from car drivers, the minority (17.7%)
from car passengers.
Table 3 shows the number of observation for different car mode /user group
combinations in the final sample.
Table 3. Number of SP-choice for different combination of user groups and evaluated car modes
Car mode the VTTS is evaluated
on
Those car
drivers
routed to
CE2d and
have car
passengers
as
alternative
All
remaining
current car
users in
the sample
Those car
passengers
routed to
CE2d and
have car
driving as
alternative
All
remaining
current car
passengers
in the
sample
Conventional car, as car driver
(CCD)
2248
17460
327
Conventional car, as car
passengers (CCP)
1705
405
3836
Partial auton. private car (PAP)
1215
256
High auton. private car (HAP)
850
220
Full auton. private car (FAP)
347
96
Full auton. ridesharing (FAR)
1343
248
As mentioned, the allocation to different introductory texts is random, but due to a
programming mistake the shares receiving the different texts are not of equal (similar) size
as seen in Table 3. It differs also between the two sources of recruitment. In Table A1 in
the Appendix, we show that the subsamples are balanced with respect to other respondent
characteristics, indicating that there seems to be no systematic in which respondents got
which intro-text. We view the allocation to introductory text therefore as exogeneous like it
was intended from the design.
6 Econometric model
We are applying the modeling approach proposed by Fosgerau et al. (2010). The binary
variable is the outcome of a choice experiment in which a respondent n is choosing
between a faster but more expensive and a cheaper but slower route alternative. The choice
experiment is repeated T times for each respondent. is defined to be of value zero if
respondent n chooses the cheaper alternative in choice task t and of value one if the
respondent chooses the faster alternative. Given the above, the decision rules is assumed to
be:
(14) iff
where is the offered trade-off (“bid”) defined as the ratio of the absolute difference of
the cost and the time attribute:
(15)
.
The error terms are assumed to be independently and identically (iid) logistic
distributed random variables with mean value of zero and variance of
3/
2
. This yields
the following cumulative distribution function in convenient closed form:
(16)
.
Rewriting the decision rule (14) we get the following probability of accepting the trade-off
(choosing the faster alternative):
(17)
The magnitude of scale parameter µ, which will be estimated from the data, will indicate
“how deterministic” the choices are. When µ approaches zero, the choices will be purely
random, resulting in a probability of 0.5 for both route alternatives.
VTTS and the bid are log transformed which restricts the VTTS to take positive values.
This is in line with the standard assumption about willingness to pay values and is also
reasonable in the VTTS context (as discussed in section 3). We further parameterize the log
of VTTS with explanatory variables.
(18)
is an (additional) error term that is assumed normally distributed with mean value zero
and variance
2
. Other than , is constant for all choices of a given respondent n.
represents therefore unobserved taste heterogeneity across respondents yielding a
lognormal distribution assumption of VTTS.
The vector x includes variables from the questionnaire that are a-prior expected to have an
effect on the VTTS. The vector includes variables describing the respondents or the
context of the reference trip, as well as the recruitment method and artifacts of the
experimental design. The latter include the choice quadrants and the size of the time saving
( : “delta T”). An overview with descriptive statistics is provided in Table 4.
Table 4: Descriptive statistics of variable in X-vector
Average values [min-
max]
Unit
Current car
drivers
Current car
passengers
Variables describing the respondents (vary over n)
Age
Amount
45.95 [16-87]
41.741 [15-84]
Age_square
Amount
2.336 [0.256-7.569]
2.045 [0.225-7.056]
Male
Dummy
0.606 [0-1]
0.29 [0-1]
Log of personal
income*
NOK per year
11.189 [0-14.447]
10.017 [0-14.447]
Income missing
Dummy
0.144 [0-1]
0.209 [0-1]
Variables describing context of the reference trip (vary over n)
Log of reference cost
NOK per trip
4.228 [0-10.26]
4.086 [0-13.956]
Log of trip distance
Kilometer
3.266 [0-8.294]
3.668 [0-8.517]
Business trips
Dummy
0.108 [0-1]
0.086 [0-1]
Commuting trip
Dummy
0.21 [0-1]
0.09 [0-1]
Leisure trip
Dummy
0.682 [0-1]
0.824 [0-1]
Variables describing the choice task (vary over n and t)
WTP-choice
Dummy
0.26 [0-1]
0.257 [0-1]
WTP-choice
Dummy
0.245 [0-1]
0.248 [0-1]
EL-choice
Dummy
0.249 [0-1]
0.248 [0-1]
EG-choice
Dummy
0.247 [0-1]
0.246 [0-1]
Log of Delta T
Minutes
2.181 [0-8.921]
2.487 [0-8.897]
Variables for recruiting (vary over n)
Email register
Dummy
0.561 [0-1]
0.49 [0-1]
Internet panel
Dummy
0.396 [0-1]
0.468 [0-1]
*including zero values for missing income.
Let’s turn to the z-vector in equation (18). The z’s are dummies identifying evaluated
transport mode and the user group. As the allocation to the different types of autonomous
vehicle is exogenous in our survey, we do not define user groups by the alternative mode
except for those respondents being routed in CE2d. Only there the choice/availability of
alternative modes (car driver and car passengers) is potentially endogenous such that one
needs to control for self-selection.
For identification, one of the coefficients has to be normalized. Let , i.e. the
user group of car drivers, evaluated at the mode car driving, be zero. For this user group:
(19)
.
And for all other combinations:
(20)
Dividing (20) by (19) and after observing the data we get:
(21)
.
Hence, the exponential for the parameters of the coefficients is the relative VTTS
compared to the VTTS of car driver group in mode car driving.
Similarly, we can identify the relative VTTS between any to combination, e.g. between car
passengers VTTS in conventional cars and in fully automated private cars:
(22)
(22) can be seen as the “relative” version of the mode effect between conventual cars and
fully automated private car (for user group car passengers). If , the
VTTS ratio would be greater than 1 indicating the car passengers experience the time in full
automated cars as more pleasant/useful/comfortable than in conventional cars. This
corresponds to the classical mode effect in absolute numbers (as stated in equation 3) being
positive (
7 Estimation model
Parameter inference is made with Biogeme (Bierlaire 2009) based on the method of
maximum log-likelihood. For the mixed logit model like the one described in the above
section, the likelihood function is simulated. For all model versions (also those used for
hypothesis testing in section 8), we use 500 Halton draws to simulate the random
parameter.
Table 5 shows the estimation results for models where none of the delta parameters is
constrained (besides the normalized group being CCD_CD, i.e. the VTTS in conventional
cars for current car drivers). We report two model versions, one without socioeconomic
variables and with the inclusion of age, gender and income. From Table 4 in the previous
section, we know that socioeconomic variables differ between car drivers and car
passengers. The comparison of the two models is therefore illustrative regarding the
relative differences between VTTS for car drivers and car passenger. For the mode effects,
which are the basis for hypotheses testing, the choice of background variables is expected
to be of minor importance.
Table 5: Estimation results
Model 1: without
sociodemographic variables
Model 2: with
sociodemographic
variables
Nr. of parameters
25
30
Nr. of observations
30556
30556
Nr. of respondents
3111
3111
Null-LL
-21179.805
-21179.805
Final-LL
-13490.459
-13452.785
Adjusted rho-squ
0.362
0.363
Value
Robust t-stats
Value
Robust t-
stats
2.26
26.35
-1.93
-3.57
σ
1.1
40.3
1.08
39.99
-0.00486
-0.49
-0.0146
-0.14
0.064
1.27
0.34
7.65
4.41
7.58
0.3
6.99
0.295
6.96
0.0141
0.32
0.00939
0.22
0.416
4.98
0.351
4.24
0.104
1.73
0.0234
0.39
0
Normalized
0
Normalized
0.639
15.49
0.639
15.51
0
Normalized
0
Normalized
0.282
7.7
0.283
7.72
0.35
9.73
0.351
9.75
0.061
2.16
0.0632
2.27
-0.13
-2.69
-0.0696
-1.43
0
Normalized
0
Normalized
0
Normalized
0
Normalized
0.115
1.38
0.126
1.53
-0.713
-3.35
-0.65
-3.1
0.0835
0.89
0.0875
0.94
-0.17
-2.26
-0.0802
-1.06
-0.251
-1.29
-0.203
-1.01
-0.204
-2.72
-0.204
-2.71
0.168
0.62
0.248
0.92
-0.157
-1.53
-0.166
-1.62
-0.47
-2.05
-0.391
-1.71
-0.349
-2.02
-0.363
-2.1
-0.944
-3.32
-0.807
-2.9
-0.237
-2.72
-0.236
-2.72
-0.242
-1.23
-0.122
-0.62
µ
1.43
14.35*
1.43
14.37*
* t-stat for mu is against value 1.
Although not the focus of the paper, we (briefly) discuss the effects of background and
design variables on the VTTS.
Based on the second model, we see that income has by far the most prominent effect on
VTTS among the socioeconomic variables. The other variables (age and gender) are not
significant (after controlling for income). The income-elasticity is 0.34 which in the lower
end of the expected range
6
.
Reference cost has a significant positive effect on VTTS, i.e. the more expensive the trip
the higher the VTTS. Note that costs for car travel is highly correlated with trip distance.
This might be the very reason why the distance-elasticity is not statistically different from
zero.
From the three trip purposes, business trips have – as expected - the highest VTTS (a
35.1% increase from leisure trips, even 41.6% without controlling for socioeconomic
variables). Somewhat surprisingly, the effect of commuting trips (compared to leisure trips)
is rather weak and not statistically significant (without controlling for socioeconomic
variables the t-stat (1.71) is decent though).
6
In the 2009 value of time study the income elasticity varied between 0.248 and 0.617 across transport
modes, with a 0.432 for short-distance car trips (Ramjerdi et al 2010).
The effects of dummies for the choice quadrant, we get the typical picture that WTA
choice have (much) higher VTTS compared to WTP choices, with EL and EG choice
somewhat in between.
The effects of delta T on the VTTS is found significantly positive; this is consistent with
earlier findings in the international and Norwegian literature (Börjesson and Eliasson
(2014), Ramjerdi et al 2010).
From the delta coefficients, we can calculate the relative VTTS compared to the
normalized group as shown in equation 20. Table 6 gives the relative VTTS values as
estimated by the second model that controls for socioeconomic variables.
Table 6: Deduced relative VTTS based on estimation model 2.
Car mode the VTTS is evaluated
Those car
drivers
routed to
CE2d and
have car
passengers
as
alternative
All
remaining
current car
users in the
sample
Those car
passengers
routed to
CE2d and
have car
driving as
alternative
All
remaining
current car
passengers
in the
sample
Conventional car, as car driver (CCD)
113.43 %
100.00 %
52.20 %
Conventional car, as car passengers
(CCP)
109.14 %
81.63 %
92.29 %
Partial auton. private car (PAP)
81.55 %
128.15 %
High auton. private car (HAP)
84.70 %
67.64 %
Full auton. private car (FAP)
69.56 %
44.62 %
Full auton. ridesharing (FAR)
78.98 %
88.51 %
Average AV
78.70 %
82.23 %
Average AV without PAP
77.75 %
66.92 %
For current car drivers (that have car passenger as an alternative, i.e. referring to the first
column in Table 6)) we see a small mode effect between car driving and being a car
passenger. This is expected but the effect seems rather small (113.4% versus 109.1%). See
next section for the testing of the related hypothesis.
For car drivers (now the general group in the 2nd column) we see clear mode effects
between AV (in all 4 degrees) and conventional car driving. On average the VTTS in AV is
78.7% of the VTTS in conventional cars (77.8% without considering partial automation).
The order of effects within the 4 degrees seems logical with the exception that the
estimated (relative) VTTS in high automation is a bit higher than in partial automation (we
expected the opposite here as one can perform more activities in high automation
compared to partial automation).
For car passengers, we estimate lower VTTS in conventional cars both as driver and as
passengers. The effects are somewhat larger based on model 1 without controlling for age,
gender and income (see Table A2 in the appendix).
What is surprising is that the mode effect between car driving and being a car passenger in
conventional car is negative (higher VTTS as a car passenger). See next section for a
discussion. The results for AV differ quite a bit for current car passengers. The result for
partial automated private (PAP) cars is somewhat difficult to interpret as – as mentioned in
section 4 – the introduction text to the respondents is ambiguous on whether current car
passengers are on the driver seat and need to observe the AV in this scenario or if they
remain on the passenger seat. The high VTTS estimated for PAP would be consistent with
the former, as this would mean that the spectrum of possible activities is reduced (such that
we expect a higher VTTS). The VTTS with full automation is rather low for current car
passengers.
8 Hypothesis testing and discussion
In this section, we formally test the hypothesis outlined in section 3. All hypotheses are
investigated with Likelihood ratio (LR) test as typically done in non-linear models (see e.g.
Fosgerau et al 2010). For the LR-tests we run constrained model versions and compare
them to the unconstrained mode version (UCM) being model 2 from the previous section
(Table 5).
Table 7 summaries the test results for hypothesis 1. We expected that the VTTS as a car
passenger was lower than as a car driver. From the estimation results (Table 6 and 7) we
found that indeed the VTTS as car passenger was relatively smaller for current car drivers
(with car passenger as an alternative). However, as now shown in Table 7, we cannot reject
the Null-hypothesis of equal size. For car passengers (with car driving as an alternative), we
find that the unexpected result shown in Table 6 regarding the mode effect between car
driving and being a car passenger is statistically significant on a 95% confidence level. This
is indeed surprisingly as it indicates that car passengers may have a lower VTTS as car
drivers in conventional cars even though they are not able to perform certain activities
while driving. A reason may be protest behavior or fatigue as those respondents may
consider it annoying to go through eight more almost identical choice cards. There is a
weak indication for that when looking at the share of respondents that always picks the
cheaper alternative. For the car passengers, the share is 17.1% in the RCE as a car driver.
This is somewhat higher as the overall share of 11.8%.
Table 7: Summary of testing of hypothesis 1 (1a and 1b)
UCM (Model 2
from section 7)
Constrained
model 1a
Constrained model
1b
Null-hypothesis
Equation 5
Equation 6
Restrictions on
parameters
None
Nr. of parameters
30
29
29
Final-LL
-13452.785
-13453.91
-13457.458
LR-test statistic against
UCM
2.25<3.84*
9.35>3.84*
Test decision
Cannot reject
Reject
Conclusion in
conjecture with the
estimated relative size
of in UCM
Car drivers have
statistically
similar VTTS as
a car driver and a
car passenger in
conventional
cars
Car drivers have
statistically different
VTTS as a car
driver and a car
passenger in
conventional cars.
Direction of effect
is unexpected.
*3.84 is the critical value at a 95% confidence with 1 degree of freedom.
Hypothesis 2 stated that we expect car driver’s VTTS to be lower in AV than in
conventional cars. The average reduction was empirically shown to be 21.3 percent points
(compare Table 6). From the formal hypothesis testing (Table 8), we see that this effect is
statistically significant as we clearly reject the null-hypothesis of equal VTTS.
The same applies for current car passenger as seen from Table 8. The result for car
passenger is, however, not so easily interpreted and for partial automation we saw – as
discussed at the end of the previous section- an increase in VTTS. Never the less the
overall effect goes in the same direction as for car drivers such that an overall decrease in
VTTS in most realistic AV scenario can be expected.
Table 8: Summary of testing of hypothesis 2 and 3
UCM (Model 2
from section 7)
Constrained
model 2
Constrained model
3
Null-hypothesis
Equation (7)
Equation (8)
Restrictions on
parameters
None
Nr. of parameters
30
26
26
Final-LL
-13452.785
-13466.026
-13458.788
LR-test statistic against
UCM
26.48>9.49*
12.00>9.49*
Test decision
Reject
Reject
Conclusion in
conjecture with the
estimated relative size
of in UCM
Car drivers have
significant lower
VTTS in AV
than in
conventional
cars
Car passengers have
different VTTS in
conventional and
automated cars but
the directions is
sensitive to the
degree of
autonomation
*9.49 is the critical value at a 95% confidence with 4 degrees of freedom.
A significant lower (future) VTTS, may have considerable impacts on the expected user
benefits of road project. This applies not just for future road projects, but also for present
ones as the emergence of AV is expected to be a reality within the typically 40 years’ time
horizon of cost-benefit analysis. Other things equal, fewer road projects will we
socioeconomically beneficial.
On the other hand, our results also indicate that the sometime proposed “zero-VTTS”
scenario is very unlikely. Also in AV do people seem to have significant willingness-to-pay
for travel time reductions.
Turning to hypothesis 4, which states that we expected lower VTTS with increasing
autonomation in private AV for car drivers. From the formal hypothesis tests (Table 9) we
see that we only can reject the null-hypothesis of equal VTTS between full and partial
automation (
in Table 9).
7
From Table 6 we deduce that he relative reduction from the
VTTS in partial AV to full AV is about 14.7% ((69.56 %/81.55 %)-1).
From a methodological point of view, we think that there are ways to improve the
communication of different degrees of autonomation to the respondents. We used rather
short introduction text, without any illustrations. We also do not know how many
respondents read and processed the information in the introduction text. The topic of AV
7
For what it is worth mentioning, hypothesis
and would be rejected with a 90% confidence level
was not a central element of the Norwegian Valuation study and we suggest more
dedicated studies on the VTTS of (different degrees of) vehicle autonomation. We
hypothesis a correlation between the reduction in VTTS and the degree to which
respondents have understood the scenarios. Unfortunately, we have no data to test this; we
suggest that future studies investigate this relationship.
Table 9: Summary of testing of hypothesis 4.
UCM (Model
2 from
section 7)
Constrained
model 4a
Constrained
model 4b
Constrained
model 4c
Null-hypothesis
Equation 9
Equation 10
Equation 11
Restrictions on
parameters
None
Nr. of parameters
30
29
29
29
Final-LL
-13452.785
-13452.914
-13454.95
-13454.448
LR-test statistic
against UCM
0.258<3.84*
4.33>3.84*
3.33<3.84*
Test decision
Cannot reject
Reject
Cannot reject
Conclusion in
conjecture with
the estimated
relative size of
in UCM
Car drivers’
VTTS in high
and partial
automated
cars is
statistically
similar
Car drivers’
VTTS in full and
is significantly
lower than in
partial
autonomation
Car drivers’
VTTS in high
and full
automated cars
is statistically
similar
*3.84 is the critical value at a 95% confidence with 1 degree of freedom.
Finally, we look at hypothesis 5 that proposed that the VTTS in shared full automated cars
is higher than in private full automated cars. The results presented in Table 10 show that
the identified differences from the estimation models are significant for current car
passengers (
can be rejected) but not for car drivers (
cannot be rejected).
Table 10: Summary of testing of hypothesis 5 (5a and 5b)
UCM (Model 2
from section 7)
Constrained model 5a
Constrained model
5b
Null-hypothesis
Equation 12
Equation 13
Restrictions on
parameters
None
Nr. of parameters
30
29
29
Final-LL
-13452.785
-13453.179
-13455.171
LR-test statistic
against UCM
0.788<3.84*
4.77 >3.84*
Test decision
Cannot reject
Reject
Conclusion in
conjecture with
the estimated
relative size of
in UCM
No significant different
in VTTS between
private and share full
automated cars for
current car drivers
Current car
passengers have
significant higher
VTTS in shared full
automated cars
compared to private
full automated cars
*3.84 is the critical value at a 95% confidence with 1 degree of freedom.
9 Conclusion
To our knowledge is this the first paper that uses route choice experiments to estimate
VTTS in scenarios of vehicle automation. The methodical approach was shown to give
some reasonable and significant results.
Car drivers’ VTTS in AVs is around 21.2% lower than in conventional cars. The reduction
is significant and expected given the larger spectrum of secondary activities that can be
performed while sitting in the car. We find that respondents are somewhat sensitive to the
degrees of automation and estimate the lowest VTTS for full autonomation in private cars
(both for current car drivers and car passengers). We also find that one has a somewhat
higher VTTS in shared AV, however this results is statistically significant only for car
passengers.
References
Bahamonde-Birke, F. J., B. Kickhöfer, D. Heinrichs, and T. Kuhnimhof (2016). A systemic
view on autonomous vehicles: Policy aspects for a sustainable transportation planning.
Tech. rep. German Aerospace Center (DLR), http://elib.dlr.de/108647/
Becker, G. (1965). A Theory of the Allocation of Time. The Economic Journal, 75: 493-517.
Bierlaire, M. (2003). BIOGEME: A free package for the estimation of discrete choice
models. Proceedings of the 3rd Swiss Transportation Research Conference, Ascona,
Switzerland.
Börjesson, M., Eliasson J. (2014). Experiences from the Swedish value of time study.
Transportation Research Part A: Policy and Practice. 59, 144-158.
DeLooff E., Correia G., Van Cranenburgh S., Snelder M., Van Arem B. (2018) Potential
changes in value of travel time as a result of vehicle automation: a case study in the
Netherlands, Submitted for Presentation in the 97th Annual Meeting of the Transportation
Research Board
DeSerpa, A. (1971). A Theory of the Economics of Time. The Economic Journal, 81: 828-846
Flügel S. (2014) Accounting for user type and mode effects on the value of travel time
savings in project appraisal: Opportunities and challenges, Research in Transportation
Economics, 2014, 47:50-60
Flügel S., Fearnley N., Tonheim C. (2017) Unge menn mest positive til bruk av førerløse
biler. Samferdsel.no (in Norwegian) https://samferdsel.toi.no/forskning/unge-menn-mest-
positive-til-bruk-av-forerlose-biler-article33888-2205.html
Fosgerau, M., Hjorth, K. & Lyk-Jensen, S. V. (2010). Between-mode-differences in the
value of travel time: Self-selection or strategic behaviour? Transportation Research Part D:
Transport and Environment, 15 (7): 370-381.
Fosgerau (2019) Automation and the value of time in passenger transport, International
Transport Forum Discussion Paper 2019
Gasser, T. M., & Westhoff, D. (2012). BASt-study: Definitions of automation and legal
issues in Germany. Presentation at the 2012 Road Vehicle Automation Workshop, Irvine,
CA.
Ho, C.Q., Mulley C.,Shiftan Y., Hensher A.D. (2015). "Value of travel time savings for
multiple occupant car: evidence from a group-based modelling approach." Proceedings to
Australasian Transport Research Forum 2015
Jara-Díaz, S. R. & Guevara, C. A. (2003). Behind the subjective value of travel time - The
perception of work, leisure, and travel from a joint mode choice activity model. Journal of
Transport Economics and Policy, 37, part 1: 29-46.
Kyriakidis M., Happee R., de Winter J.C.F. (2015) Public opinion on automated driving:
Results of an international questionnaire among 5000 respondents, Transportation Research
Part F: Traffic Psychology and Behaviour (32), p-127-140,
https://doi.org/10.1016/j.trf.2015.04.014
Oort, C. (1969). The evaluation of travelling time. Journal of Economics and Policy, 3 (3): 279-
286.
Ramjerdi, F., Flügel, S., Samstad, H. & Killi, M. (2010). Den norske verdsettingsstudien -
Tid. Oslo: Institute of Transport Economics. (in Norwegian)
Steck. F., Kolarova V., Bahamonde-Birke F., Trommer S., Lenz B. (2018) How
Autonomous Driving May Affect the Value of Travel Time Savings for Commuting
Transportation Research Record 1 –10, 2018
Yap, M. D., Correia G., van Arem B. (2016) Preferences of travellers for using automated
vehicles as last mile public transport of multimodal train trips. Transportation Research Part A:
Policy and Practice, No. 94, 2016, pp. 1–16.
Appendix
Table A1: Descriptive statistics for 4 subgroups
Partial, private
High, private
Full, private
Full, rideshare
Mea
n
95.0
%
Lower
CL for
Mean
95.0
%
Upper
CL for
Mean
Mea
n
95.0
%
Lower
CL for
Mean
95.0
%
Upper
CL for
Mean
Mea
n
95.0
%
Lower
CL for
Mean
95.0
%
Upper
CL for
Mean
Mea
n
95.0
%
Lower
CL for
Mean
95.0
%
Upper
CL for
Mean
alder
46.6
1
45.83
47.39
47.0
2
46.17
47.87
43.1
8
41.90
44.47
47.9
2
47.12
48.73
age_squ_over_100
0
2.40
2.33
2.48
2.41
2.33
2.49
2.05
1.94
2.17
2.57
2.49
2.64
kjonn
0.54
0.52
0.57
0.57
0.54
0.60
0.49
0.44
0.53
0.56
0.54
0.59
ln_inntekt
13.0
4
13.01
13.08
13.1
3
13.09
13.16
13.0
2
12.94
13.09
12.9
8
12.95
13.02
d_missinntekt
0.14
0.12
0.16
0.12
0.10
0.14
0.13
0.10
0.16
0.10
0.09
0.12
ln_BaseCost
4.34
4.25
4.44
4.76
4.65
4.87
4.59
4.44
4.75
3.99
3.91
4.08
ln_Dist
3.43
3.34
3.52
3.84
3.73
3.95
3.78
3.64
3.92
3.09
3.01
3.17
Treise_d
0.10
0.09
0.12
0.10
0.08
0.11
0.13
0.10
0.16
0.11
0.10
0.13
Areise_d
0.18
0.16
0.20
0.18
0.16
0.20
0.18
0.14
0.21
0.21
0.19
0.23
Freise_d
0.72
0.69
0.74
0.73
0.70
0.75
0.70
0.65
0.74
0.68
0.66
0.70
ln_delta_T
2.30
2.23
2.36
2.60
2.52
2.68
2.50
2.40
2.61
2.05
1.99
2.11
d_WTP
0.25
0.23
0.28
0.25
0.22
0.28
0.25
0.21
0.29
0.25
0.23
0.27
d_WTA
0.25
0.23
0.27
0.25
0.23
0.28
0.25
0.21
0.29
0.25
0.23
0.27
d_EG
0.24
0.22
0.26
0.24
0.22
0.27
0.24
0.20
0.28
0.25
0.22
0.27
d_EL
0.26
0.24
0.28
0.25
0.23
0.28
0.25
0.21
0.29
0.25
0.23
0.27
d_panel
0.38
0.35
0.40
0.48
0.45
0.51
0.45
0.40
0.49
0.44
0.42
0.47
d_email
0.62
0.60
0.65
0.52
0.49
0.55
0.55
0.51
0.60
0.56
0.53
0.58
Table A2: Deduced relative VTTS based on estimation model 1.
Car mode the VTTS is evaluated
Those car
drivers
routed to
CE2d and
have car
passengers
as
alternative
All
remaining
current car
users in the
sample
Those car
passengers
routed to
CE2d and
have car
driving
All
remaining
current car
passengers
in the
sample
Conventional car, as car driver (CCD)
112.19 %
100.00 %
49.02 %
Conventional car, as car passengers
(CCP)
108.71 %
77.80 %
84.37 %
Partial autonomous private car (PAP)
81.55 %
118.29 %
Highly autonomous private car (HAP)
85.47 %
62.50 %
Fully autonomous private car (FAP)
70.54 %
38.91 %
Fully autonomous ridesharing (FAR)
78.90 %
78.51 %
Average AV
79.11 %
74.55 %
Average AV without PAP
78.30 %
59.97 %