CHARGING BEHAVIOUR
4.3 Response to Dynamic Pricing
4.3.3 Practical applications of EUT and non-EUT
EUT and non-EUT methods have been widely adopted in transportation research, especially during the last decade. Examples of their applications can be found in equilibrium modelling, valuation of travel time or attitudes towards travel time uncertainty and experience/learning in travel choices (Kemel and Paraschiv, 2013). The majority of these applications are based on stated preference surveys because of the level of control that they offer to the analyst. In experimental economics it is common to adopt an intermediate approach between stated and revealed preferences, i.e. incentivised laboratory experiments. Nevertheless, transport choices are more complex since they do not involve only money, but other characteristics as well (e.g.
time or comfort).
For risky choices in transport applications, EUT was often embedded with random utility models (RUM) creating an ad hoc modelling approach. Integrating these two theories provides the possibility to represent simultaneously two distinct sources of uncertainty: the uncertainty
of the decision-maker when presented with various possible outcomes (EUT) and the uncertainty of the researcher when observing the choice process of the decision-makers (RUM). The integrated RUM-EUT model is formed by adding an RUM-related error term to the value term of the EUT model of equation 4.41:
π’(π π) = β πππ[π£(π ππ)
πΎ
π=1
+ ππ] (4.47)
Such integration is governed by some important conceptual shortcomings that have been addressed by the framework of Liu and Polak (2007). This framework allows the explicit modelling of attitudes towards risk independent of the conventional RUM-like tastes.
It is assumed that π½π = {π½ππ;1 β€ π β€ π } is a set of taste parameters associated with the vector of observable attributes π΄ = {ππ1 β€ π β€ π} that affect the choice behaviour. Taste parameters are individual-specific and they are characteristic of a riskless choice. Each individual π attaches a scalar value π£πππ to the kth outcome of prospect π π which is a function of π½π, i.e. π£πππ = π(π ππ, π½π). The scalar value for prospect π π (π’ππ) is a function of π£πππ, the probabilities associated with prospect outcomes ππ and a set of extra parameters ππ = {πππ; 1 β€ π β€ π} that are representative of decision-making under risk. Hence, it can be denoted as π’ππ = π(π£π1π, β¦ , π£ππΎππ1π, β¦ , ππΎπ, ππ).
Allowing some relaxation in EUT, each outcome π ππ instead of being evaluated based on its riskless value π£πππ can be evaluated based on a non-linear function z(π£πππ, ππ). When this function is concave, then the expected utility is lower than the expected value and hence the individual is risk-averse. On the other hand, when this function is convex the individual is risk prone. A typical non-linear transformation of the value function is when the individuals are characterised by constant absolute risk aversion (CARA), i.e. the risk attitudes for a particular prospect are not affected by the value of the outcomes involved in this prospect. The functional form of CARA is π§(π₯) = (1 β πβππ₯)/π where a positive π expresses risk proneness while a negative π expresses risk aversion.
Integrating EUT and RUM, the scalar value for prospect π π can now be transformed to:
π’ππ = β π§(π£πππ, ππ)πππ
πΎ
π=1
+ πππ (4.48)
where πππ is the unobservable component48 of the utility associated with prospect π π and π£πππ = π€πππ + ππππ is the combination of observable and unobservable components of the value function associated with outcome k.
In their majority, individuals are risk-averse and in a dynamic pricing context, they tend to choose a certain price over an uncertain one (Bonsall and Shires, 2005). Moreover, it has been found that the upper end of the price distribution disproportionally affects their choice. For example, EV drivers would avoid charging their vehicle during a specific period if there was a possibility of a very high cost.
The assessment of state-dependent prices, according to empirical evidence, strongly depends on individualsβ expectations (Lindsey, 2011). The judgement of dynamic prices is affected by historical prices as well as by prices charged for similar services or under similar circumstances. These reference-dependent effects are likely to decay for increasing experience with the price mechanism, as individuals become more familiar with gains and losses.
Potentially, this is the reason dynamic pricing has become acceptable for airline tickets and other travel-related services (e.g. dynamic tolls).
Prospect theory is suitable for modelling the response to price signals because it has been observed that choices depend on how the individual perceives the monetary transaction: in a positive or negative way. For example, paying for out-of-home charging might be considered as a loss, due to the fact that this service is free for most of the respondents at the time being.
4.3.4 Empirical estimation
For the booking game of the EV-PLACE survey, as it was presented in Chapter 3, respondents had to choose between one certain and one risky alternative according to the activities and the timings of the home-based tour that they have selected earlier. The background characteristics like the location of the charging post were similar to the charging game. Also, the charging attributes, i.e. charging duration, CISDE and walking time to the destination, were randomly selected from the design levels of the previous SP exercise and were fixed across the choice situations.
Respondents are presented with a deterministic price (βBOOK NOWβ, the safe option) and with a random price (βBOOK LATERβ, the risky choice). Specifically, πΆπ is the deterministic price of the βbooking nowβ choice, πΆπ· is the decreased potential price of the risky choice while
48 Note that the error term πππ applies now to the whole prospect and not to the expected term as in equation 4.47.
πΆπΌ is the increased potential price of the risky choice. If the integrated RUM-EUT is treated as the basic model specification, then the utility functions for the two booking alternatives, for the riskless form of 4.48, are given by:
π’ππππ = π΄ππΆπππ+ π½πΆπΆπ+ ππ (4.49)
π’ππΏπ΄ππΈπ = π΄ππΆπΏπ΄ππΈπ + π½πΆ[ππΌπΆπΌ+(1 βππΌ)πΆπ·] + ππ (4.50) where ππΌ is the probability for a future increase in price, (1 β ππΌ) is the probability for a future decrease in price, ππ is the error associated with the analystβs observations, π΄ππΆπππ and π΄ππΆπΏπ΄ππΈπ are the alternative specific constants and π½πΆ is the sensitivity to price.
The estimation of the model was carried out with BIOGEME 2.2 (Bierlaire, 2003). Like for the charging choice model in subsection 4.2.4, a scale parameter is estimated in order to reduce the variance of the error term for the Panelbase respondents. The parameter estimates and the goodness-of-fit of the model are presented in Table 4.11.
The results of the RUM β EUT model are quite intuitive, with a positive constant for the βbook nowβ option (i.e. an implicit preference for the non-risky choice, indicating a tendency towards myopic behaviour), and a statistically significant negative coefficient for the charging price, which is very close to the estimation from the charging game (π½πΆπ = β0.918 in Table 4.4).
Table 4.11: RUM β EUT model, booking game β base specification
Variables Coefficient Std error t-test p-value
ASCNOW 1.18 0.139 8.45 0.00**
In order to capture systematic heterogeneity and the effect of socio-demographics and other attributes on the booking choice, equation 4.49 is transformed to:
π’ππππ = π΄ππΆπππ+ π½πΆπΆπ+π½πβ²πΏ + π (4.51) where πΏ is the vector of personal attributes and π½πβ² is the associated vector of parameters to be estimated. Several specifications have been tested and the majority of the variables that enter the utility function are similar with the MNL model for the charging game. Additional (or modified) explanatory variables are presented below:
ο· Age group (Over 60, Less than 60)
ο· Travel profile frequency (Undertake the selected travel profile every day, undertake the selected travel profile less frequently)
ο· Schedule Flexibility (Combined Likert scale values of the indicators presented in subsection 4.2.4)
ο· Charging frequency (Charging the EV more than once a day, Charging the EV less than once a day, Non-EV drivers)
ο· Daily mileage with EV (Driving more than 40 miles a day with EV, Driving less than 40 miles a day with EV, Non-EV drivers)
The estimates for the full specification, after accounting for systematic heterogeneity, are shown in Table 4.12.
Following the framework of Liu and Polak (2007), the CARA functional form has been adopted for the non-linear transformation of the value function49. The non-linear formulation allows the researchers to investigate the decision makerβs attitude towards risk in the dynamic pricing environment. The utility function for the risky alternative is now:
π’ππΏπ΄ππΈπ = π΄ππΆπΏπ΄ππΈπ + π½πΆ [ππΌ(1 β πβππΆπΌ)/π +(1 βππΌ)(1 β πβππΆπ·)/π)] + ππ (4.52) The estimation results from the non-linear transformation are presented along with the riskless value function in Table 4.12.
The goodness-of-fit for the full specification of the linear model is significantly improved compared to the base specification (πΜ =0.232 vs πΜ =0.178). According to the estimated parameters, older individuals are more likely to choose the safe option compared to younger individuals. Similar are the findings for employed individuals. Under a different risky context, Daina (2014) has also found that older groups and people with full-time employment
49 The constant relative risk-aversion (CRRA) transformation has been also tested with the empirical data but the
demonstrate a higher risk aversion, reflected by their increased sensitivity to a latent construct for range anxiety.
On the other hand, those that have children and a higher education level tend to exhibit a more strategic behaviour. Work-based tours are associated with a more conservative response to dynamic pricing, i.e. an increased preference for the βbooking nowβ option. Finally, people that own or lease an EV are more risk-prone and are willing to wait for a better price.
Table 4.12: RUM β EUT model, booking game β full specification accounting for systematic heterogeneity
Variables Linear function Non-linear function
Coefficient Std error Coefficient Std error
ASCNOW 1.49* 0.805 3.27* 1.94
Education: University Graduate -0.964** 0.320 -0.977** 0.323
Electric vehicle access -0.847* 0.509 -0.861* 0.511
Number of daily activities -0.272 0.229 -0.274 0.230
Number of profile searches 0.324* 0.196 0.323* 0.198
Travel profile β Every day 1.02 0.720 1.03 0.722
Travel day β Weekday 0.542 0.370 0.550 0.371
Work based tour 0.924** 0.387 0.929** 0.388
Schedule flexibility -0.0376 0.0320 -0.0377 0.0321
Charge EV more than once a day 0.908* 0.548 0.920* 0.552
Charging EV cost β free 5.04** 2.39 5.10** 2.38
Driving EV more than a year -1.01** 0.323 -1.02** 0.324
EV loyal enthusiast 0.792** 0.322 0.800** 0.324
Daily mileage with EV β more than 40 miles -1.07** 0.414 -1.09** 0.417
Scale for recruitment channel (πΌ) 0.416** 0.0771 0.415** 0.0762
Risk attitude parameter (πΆ) - - 0.155 0.128
EV drivers who stated that they charge their vehicle more than once a day demonstrated a myopic behaviour, which could be attributed to the planning burden associated with monitoring dynamic prices every time they charge. Furthermore, individuals that have been
possibly because they have a lower willingness to take the risk for an increased charging price.
Experienced EV drivers have a lower likelihood of being myopic while those that are labelled as βEV loyal enthusiastsβ have a lower likelihood of being strategic. Finally, individuals that have been regularly driving long distances with their electric vehicle prefer the risky choice.
A potential explanation for their behaviour is that they are more familiar with risky situations, due to the fact that they repeatedly strain the limits of their battery range.
The goodness-of-fit for the non-linear transformation is slightly lower than the linear function and there is no significant difference in the parameter estimates. As with the case of mixed logit in 4.2.4.3, there could be an issue of overfitting, and cross-validation is required in order to identify it. The alpha parameter is not significant but this can be explained by the fact that risk aversion is captured by the coefficient for the alternative specific constant of the βbooking nowβ option, which is positive and significant.
The non-EUT approaches presented in subsection 4.3.2 are also applied here, in order to identify misconceptions, biases and errors in the choice process of individuals.
First, the utility for the risky alternative is transformed based on the RDEU model as follows:
π’ππΏπ΄ππΈπ = π΄ππΆπΏπ΄ππΈπ + π½πΆ [π€(ππΌ)(1 β πβππΆπΌ)/π + (1 β π€(ππΌ))(1 β πβππΆπ·)/π)] + ππ (4.53) where w(.) is a function that reflects the individual weights towards risky outcomes and the outcomes are ranked in an increasing preference order (i.e. the increased price is ranked first and the decreased price is ranked second). This function is given by:
π€(π ππ) = {π(π1π, π2π) β π(π2π) ππ π = 1
π(π2π) ππ π = 2 (4.54) and π(. ) is the increasing weighting function of probability πππ with π(0) = 0 and π(1) = 1.
Also, π(π1π, π2π) is the weight associated with obtaining outcome 1 or better than 1.
The weighting function w(.) that is employed is this of equation 4.43, which results into an inverse-S shaped curve. The results from the estimation of the linear value function and the non-linear transformation under the RDEU model are presented in Table 4.13.
The model fit for the non-EUT models has not improved relative to the previous specifications.
The signs and magnitudes of the estimates remain similar to the EUT model and like before, the parameter for the attitude towards risk is insignificant.
Table 4.13: RUM β RDEU model, booking game - full specification accounting for systematic heterogeneity
Variables Linear function Non-linear function
Weighting function (πππ) = (πππ)πΎ
((πππ)πΎ+ (1 β (πππ)πΎ))1πΎ
) (πππ) = (πππ)πΎ ((πππ)πΎ+ (1 β (πππ)πΎ))1πΎ
)
Coefficient Std error Coefficient Std error
ASCNOW 3.07* 1.68 6.23* 3.49
Education: University Graduate -0.982** 0.325 -0.999** 0.329
Electric vehicle access -0.911* 0.534 -0.931* 0.537
Number of daily activities -0.272 0.233 -0.274 0.235
Number of profile searches 0.330* 0.198 0.330* 0.200
Travel profile β Every day 1.02 0.739 1.04 0.743
Travel day β Weekday 0.537 0.382 0.547 0.383
Work-based tour 0.963** 0.401 0.972** 0.402
Schedule flexibility -0.0380 0.0327 -0.0382 0.0329
Charge EV more than once a day 0.885 0.555 0.898 0.560
Charging EV cost β free 5.18** 2.51 5.25** 2.51
Driving EV more than a year -1.02** 0.327 -1.03** 0.330
EV loyal enthusiast 0.802** 0.327 0.813** 0.330
Daily mileage with EV β more than 40 miles -1.06** 0.420 -1.09** 0.424
Scale for recruitment channel (πΌ) 0.396** 0.0758 0.395** 0.0742
Risk attitude parameter (πΆ) - - 0.198 0.140
The parameter πΎ is statistically significant for both specifications. As a result, it contains information about the individualsβ perceptions of the probabilities for risky outcomes. The value of πΎ is close to 1 so the distortion of the objective probabilities is small. Nevertheless, instead of the commonly encountered inverse S-shape, this weighting function has the opposite effect. This means that individuals slightly underweight low probabilities and overweight high probabilities. The crossover point for the inverse-S curve is around 0.5. The relationship between probabilities and weighted probabilities for the linear value function (it is almost the
Figure 4.10: Objective and subjective probabilities of the risky outcomes
Since the probabilities of future prices are based on an orthogonal design, the aggregated probabilities for both outcomes throughout the choice experiment are 0.5 and, as a result, their decision weights are equal. The results suggest that when the probability of an increased price is small (i.e. 0.2 and w(0.2)<0.2 and 1-w(0.2)>0.8) then an increased price weights less than proportionally compared to the objective probability, reflecting optimism for the individuals.
However, this optimism is small because πΎ is close to 1. On the other hand, when the probability of a decreased price is small, following the same logic, individuals show some pessimism and tend to be risk-averse.
Following the Prospect Theory approach, the utility function for the βbooking laterβ can be expressed as follows:
π’ππΏπ΄ππΈπ = π΄ππΆπΏπ΄ππΈπ + π½πΆ(ππππ)(1 β ππΌ)(πΆπ ππβ πΆπ·)π+ π½πΆ(πππ π )ππΌ(πΆπΌβ πΆπ ππ)π½+ ππ (4.55) where πΆπ ππ is the reference price, which due to the nature of the problem is assumed to be equal to the price for the safe choice, i.e. Β£2.5050. The cost coefficient is now divided into gain π½πΆ(ππππ) and loss π½πΆ(πππ π ) based on the relative location of the outcome with respect to the reference price πΆπ ππ. Finally, the parameters π and π½ reflect the degree of diminishing sensitivity.
50 It has to be noted here that this price level might be non-representative of the existing recharging cost for EV drivers. In this case, it wouldnβt coincide with the reference price from the revealed preferences of the individuals.
Nevertheless, it is undoubtedly the βreference pointβ for the hypothetical choice of the booking game, based on
Two PT models were estimated: one where the π and π½ parameters are fixed to one in order to capture only reference dependence, and one that allows the estimation of diminishing sensitivity. The estimation results are presented in Table 4.14.
Table 4.14: RUM β PT model, booking game - full specification accounting for systematic heterogeneity
Variables Reference dependence Diminishing sensitivity
Coefficient Std error Coefficient Std error
ASCNOW 2.14** 0.927 2.76** 1.03
Education: University Graduate -0.981** 0.324 -1.00** 0.329
Electric vehicle access -0.866* 0.512 -0.922* 0.533
Number of daily activities -0.274 0.231 -0.276 0.235
Number of profile searches 0.323* 0.198 0.330* 0.200
Travel profile β Every day 1.04 0.724 1.04 0.743
Travel day β Weekday 0.553 0.372 0.552 0.383
Work-based tour 0.933** 0.388 0.971 ** 0.401
Schedule flexibility -0.0379 0.0322 -0.0386 0.0329
Charge EV more than once a day 0.923* 0.553 0.906 0.560
Charging EV cost β free 5.11** 2.38 5.23** 2.50
Driving EV more than a year -1.02** 0.325 -1.04** 0.330
EV loyal enthusiast 0.804** 0.325 0.818** 0.330
Daily mileage with EV β more than 40 miles -1.09** 0.418 -1.09** 0.424
Scale for recruitment channel (πΌ) 0.414** 0.0756 0.397** 0.0738
πΆ - - 0.960** 0.249
The model fit is similar to the RDEU model and the results agree with the a priori expectations, i.e. the price coefficient is positive when the outcome is framed as a gain and negative when it is framed as a loss. While the positive and significant parameter for the constant of the βbook nowβ option indicates a risk-aversion for both models, the absolute ratio of the cost coefficients