The neighbor effect : Simulating dynamics in consumer preferences for new vehicle technologies

Transcription

1 available at ANALYSIS The neighbor effect : Simulating dynamics in consumer preferences for new vehicle technologies Paulus Mau a,1, Jimena Eyzaguirre a,1, Mark Jaccard a,, Colleen Collins-Dodd b,2, Kenneth Tiedemann a,1 a Simon Fraser University, School of Resource and Environmental Management, 8888 University Drive, Burnaby, BC, Canada V5A 1S6 b SFU Business, Simon Fraser University Vancouver, 515 West Hastings Street, Vancouver, BC, Canada V6B 5K3 ARTICLE INFO Article history: Received 28 April 2007 Received in revised form 9 April 2008 Accepted 7 May 2008 Available online 28 June 2008 Keywords: HEV HFCV Stated preferences DCM Consumer preference dynamics ABSTRACT Understanding consumer behaviour is essential in designing policies that efficiently increase the uptake of clean technologies over the long-run. Expert opinion or qualitative market analyses have tended to be the sources of this information. However, greater scrutiny on governments increasingly demands the use of reliable and credible evidence to support policy decisions. While discrete choice research and modeling techniques have been applied to estimate consumer preferences for technologies, these methods often assume static preferences. This study builds on the application of discrete choice research and modeling to capture dynamics in consumer preferences. We estimate Canadians' preferences for new vehicle technologies under different market assumptions, using responses from two national surveys focused on hybrid gas-electric vehicles and hydrogen fuel cell vehicles. The results support the relevance of a range of vehicle attributes beyond the purchase price in shaping consumer preferences towards clean vehicle technologies. They also corroborate our hypothesis that the degree of market penetration of clean vehicle technologies is an influence on people's preferences ( the neighbor effect ). Finally, our results provide behavioural parameters for the energy-economy model CIMS, which we use here to show the importance of including consumer preference dynamics when setting policies to encourage the uptake of clean technologies Elsevier B.V. All rights reserved. 1. Introduction Policymakers committed to achieving environmental goals face significant risks in deciding among policy options, and require credible information to support the design of effective policies that minimize societal costs over the long-run. In cases where long term environmental goals are at stake, information for policymakers about the potential effects of alternative policies tends to be scarce and uncertain, and policymakers often rely on energy-economic simulation models to make the best use of limited information. To be useful to policymakers, energyeconomic simulation models must provide the most realistic projections possible, based on the best available data. The model CIMS incorporates consumer behaviour in its modeling capabilities to realistically simulate policies aimed at causing profound technological change in the long-run (Jaccard et al., 2003). Over the past few years, CIMS' ability to portray consumer behaviour has improved for specific technologies and applications in Corresponding author. Tel.: ; fax: address: Jaccard@sfu.ca (M. Jaccard). 1 Tel.: ; fax: Tel.: ; fax: /$ see front matter 2008 Elsevier B.V. All rights reserved. doi: /j.ecolecon

2 505 industrial steam generation (Rivers and Jaccard, 2005), residential heating (Jaccard and Dennis, 2006) and personal urban transportation (Horne et al., 2005; Rivers and Jaccard, 2005). However, a major assumption in these works is that the way in which consumers value technologies and choose among them does not change that is, the portrayal of consumer preferences is static. The incorporation of preference dynamics in CIMS would enhance the model's behavioural realism and its ability to provide realistic market forecasts. This paper discusses the results of two parallel research projects that provide an empirical basis for the incorporation of dynamic preferences in CIMS. Our research focused on eliciting consumer preferences for hybrid gas-electric vehicles (HEVs) and hydrogen fuel cell vehicles (HFCVs), while manipulating the conditions surrounding the respondents' decision environment in order to trigger changes in these preferences. Because these are new technologies, it is not possible to estimate revealed preferences from past consumer behaviour. Instead, discrete choice surveys are needed to elicit stated preferences from questionnaires in which respondents are presented with hypothetical, but realistic, choice sets. Specifically, the HEV is an evolutionary innovation. Evolutionary technologies are new technologies that provide the same service as a mainstream technology without requiring major changes in infrastructure or a significant amount of learning by the consumer (Jackson, 2003; Christensen, 1997). The HFCV is a disruptive innovation. Disruptive innovations are new technologies possessing attributes initially unfamiliar to consumers and producers, perhaps requiring major changes to infrastructure and institutions as well, with the potential to shift market structures and induce behavioural change (Mackay and Metcalfe, 2002). In this research, we hypothesized that people's value for hybrid gas-electric and hydrogen fuel cell vehicles change as the number of people owning these types of vehicles increase; we refer to this phenomenon as the neighbor effect. In addition, we hypothesized that the neighbor effect would differ between evolutionary and disruptive technologies. By quantifying the neighbor effect for different technology types and including these preference dynamics in energy-economy simulation models, we may be able to enhance the behavioural realism of these models and thus we will be able to better inform policy makers seeking to induce technological change across the economy over the long run. The remainder of paper is structured as follows. Section 2 describes the CIMS model. Section 3 describes how data were collected and analyzed. Section 4 describes the micro-economic models built from the consumer data and how we used these models to estimate behavioural parameters and incorporate preference dynamics in CIMS. This section also discusses key findings related to Canadians' preferences for hydrogen gasoline-electric and hydrogen fuel cell vehicles. In Section 5, we comment on the implications of our findings and illustrate the added value of enhancing the realism of consumer preferences for new technologies in CIMS. 2. The CIMS model CIMS is a capital vintage model, which tracks the evolution of capital stocks over time through retirements, retrofits, and new purchases. The model simulates technological change in four inter-dependent steps: (1) calculating the costs for each energy service, such as the cost per person-kilometres-travelled for transportation; (2) retiring capital stock according to an agedependent function; (3) calculating market shares for new capital stock; and (4) simulating effects among energy intensive economic sectors by iterating between a macroeconomic component and the energy sectors (supply and demand) until an equilibrium is reached (Jaccard et al., 2003). The energy requirements in step one of the simulation period are directly influenced by a process called technology competition, which is how technologies offering similar services gain or lose new market shares relative to each other over a simulation period. CIMS allocates new market shares for each technology by simulating competition at each service node according to the logistic relationship in Eq. (1): MS j ¼ P K k¼1 v LCC j ½LCC k Š v MSj is the market share of technology j, k is the number of all alternatives, LCC denotes an aggregation of costs that consumers owning that technology will incur over its lifespan (lifecycle costs), and v is a measure of market heterogeneity. Eq. (2) summarizes the LCC of each technology as experienced by consumers and firms: r LCC j ¼ CC j T 1 ð1 þ rþ n þ MC j þ EC j þ i j ð2þ CCj is the capital cost (for example, sale price of vehicle), MCj are maintenance and operation costs exclusive of energy costs, ECj is the energy cost (for example, gasoline in the case of conventional cars), i j is the intangible cost (monetized value reflecting the fact that consumers and firms can attach real or perceived costs to one technology relative to another even though the two might provide the same service at similar financial costs), r is the private discount rate, and n is the technology's lifespan. For a competition node with only two technologies, the new market share of a technology varies logarithmically with changes in the ratio of LCC. The inverse power function, with v as a key parameter, acts to distribute the penetration of a particular technology j relative to all other technologies k at the node. A technology's intangible cost (i j ) is one of CIMS' parameters that captures consumer behaviour. Historically, CIMS' intangible costs came from a combination of literature review, judgement, and meta-analysis (Jaccard et al., 2004; Murphy, 2000; Nyboer, 1997). Recent research has focused on estimating intangible costs for selected technologies by using stated preference data from discrete choice experiments. Yet, these improvements in estimates of intangible costs are fixed inputs in CIMS. In addition to a static representation of intangible costs throughout a simulation exercise in CIMS, the model can accommodate changes in intangible costs in a dynamic function. Eq. (3) illustrates how CIMS can represent the intangible costs of technologies at any time I(t) (Rivers and Jaccard, 2005): I ðþ t ¼ I ðþ o ð3þ 1 þ ATe kt NMS ðt 1Þ ð1þ

3 506 ECOLOGICAL ECONOMICS 68 (2008) The parameters A and k are scalar parameters, whereas NMS represents the new market share of the technology from the previous simulation period. This function is dependent on consumer behaviour towards new technologies, and it can be estimated from consumer data using discrete choice models developed in this research. Using empirical data to parameterize this intangible cost function will allow CIMS to represent dynamic consumer behaviour towards new vehicle technologies as market conditions change through the simulations. In this way, policy simulations in CIMS will benefit from an added degree of realism in consumer behaviour, rather than simplifying consumer preferences as static and independent of basic market conditions. 3. Methods Our research methods for estimating empirically-based behavioural parameters for CIMS were adapted from those of Jaccard and Dennis (2006), Rivers and Jaccard (2005), and Horne et al. (2005). We built on their work by attempting to capture the evolution of preferences over the long-run, as new technologies evolve and gain market share. CIMS' behavioural parameters derive from the results of discrete choice models, which were informed by two surveys that were conducted in this study in eliciting Canadians' vehicle preferences through discrete choice experiments. Although virtually identical in design, one survey focused on eliciting preferences for hybrid gas-electric vehicles (HEVs) and the other for hydrogen fuel cell vehicles (HFCVs). To reproduce the neighbor effect, we divided both survey samples into four different treatment groups, respectively. We ascribed each group a hypothetical condition describing the current market share ratio of HEV or HFCV to conventional gasoline vehicles. The hypothetical market share ratios selected were: less than one percent (0.03 percent), five percent, ten percent, and twenty percent. For both the HEV and the HFCV studies, these treatment groups are referred to as market share groups Building discrete choice models Data collection and survey design We used choice experiments contained in two independent surveys to gather Canadians' preferences for HEVs and HFCVs, respectively. In these stated preference surveys, we presented participants with sets of hypothetical situations, and asked them to choose between two vehicle technologies: HEV and conventional gasoline in the HEV study, and HFCV and conventional gasoline in the HFCV study. All participants completed the survey online, which was designed to mitigate some of the bias related to task complexity without compromising efficiencies in data collection (Train, 2003; Fujii and Garling, 2003; DeShazo and Fermo, 2002; Hensher et al., 1999; Urban et al., 1996). To enable comparison of results between the two studies, the survey designs were almost identical, with both including (1) a choice experiment and (2) specific information reflective of hypothetical market share conditions to test for preference dynamics. Synovate, a Canadian marketing firm, procured a survey sample for the HEV and HFCV studies. At the time of the study, this firm maintained a national panel consisting of 70,000 Canadian households who were willing to take part in online surveys. Participants were sampled to match our population of interest: Canadian residents over 19 years of age who commute to work or school at least once a week, own at least one conventional gasoline vehicle, and reside in urban centres with a population greater than 250,000 people. Our assumption in targeting this demographic was that this segment would yield consumer choices informed of the costs of gasoline vehicles, the vehicle purchasing process, and the everyday hassles associated with vehicles, such as the inconvenience of fill-ups and maintenance. In both the HEV and HFCV studies, designing the choice experiment involved selecting the vehicle attributes and levels that would allow us to estimate, with some degree of confidence, the key influences in consumer choices between a conventional gasoline vehicle and the new vehicle technology. Several studies have investigated personal vehicle preferences using discrete choice experiments (Horne et al., 2005; Brownstone et al., 2000; Brownstone and Train, 1999; Ewing and Sarigöllü, 2000; Greene, 1997; Bunch et al., 1993). Our research considered a broader range of attributes than these studies, with the objective of improving the characterization of consumer behaviour. We chose attributes that allowed us to quantify aspects of consumer purchasing behaviour susceptible to policy influence, while also taking into account attributes that ranked high in influencing consumer purchase decisions from previous studies (Horne et al., 2005; Brownstone et al., 2000; Ewing and Sarigöllü, 2000; Brownstone and Train, 1999; Greene, 1997; Bunch et al., 1993). These attributes are the following: vehicle purchase price (also referred to as capital cost); fuel cost; the amount of a subsidy provided by the federal government as a rebate for purchasing a given vehicle technology; warranty coverage; cruising range (the distance one can drive between refuelling) for the HEV study; and refuelling convenience (the relative proportion of stations with proper fuel) for the HFCV study. The selection of these attributes, which dominated a consumer's decision in purchasing vehicles, allowed us to minimize the error in transgressing the IID (independent and identical distribution of error term across alternatives) assumption of the Logit model. Further discussion on this assumption is in the next section. The experimental design provided three levels for each of the six vehicle attributes being tested in this study high(2), medium(1), and low(0), as presented in Table 1. A key feature of the design was the customization of attributes based on participants' current vehicle choices, so that values for fuel cost, purchase price, and cruising range are a variation on participants' current vehicle. The purchase price attributes were calculated as a factor of the participant's vehicle cost, User CC ; the fuel costs were calculated as a factor of the participant's current fuel bill, User FC ; and the cruising range of HEVs was calculated as a factor of the cruising range of each participant's current vehicle, User CR. Specific to the HEV study, the fuel cost to the HEV choices were adjusted downwards with the adjustment factor k, a ratio of the cruising range of HEV and the conventional vehicle. For example, if the HEVs have 200% higher cruising range than the conventional vehicle, k would be 0.5, which would reduce the fuel cost by half. Tailoring the values in the choice experiment enhances

4 507 Table 1 Attribute levels in the choice experiments Attribute Vehicle Attribute levels are relative to the attributes of the participant's current vehicle technology Low (0) Medium(1) High(2) Capital cost Conventional User CC User CC 110% User CC 120% HEV and HFCV User CC 140% User CC 170% User CC 190% Fuel cost Conventional User FC User FC 110% User FC 125% HEV User FC k User FC k 110% User FC k 125% HFCV User FC User FC User FC Government subsidy Conventional None None None HEV 5% of HEV capital cost 10% of HEV capital cost 20% of HEV capital cost HFCV 5% of HFCV capital cost 10% of HFCV capital cost 20% of HFCV capital cost Cruising range Conventional User CR User CR User CR (HEV study only) HEV User CR 120% User CR 150% User CR 200% Stations with proper fuel Conventional All stations All stations All stations (HFCV study only) HFCV 1 in 20 stations 1 in 10 stations 1 in 5 stations Warranty Conventional 5 years or 100,000 km 5 years or 100,000 km 5 years or 100,000 km HEV and HFCV 5 years or 100,000 km 8 years or 130,000 km 10 years or 163,000 km User CC = Capital cost of the survey participant's current vehicle; User FC = Fuel cost of the survey participant's current vehicle; k = variable scaling ratio representing the efficiency gain in hybrid vehicles. each survey participant's ability to relate to the survey questions and make informed choices. Although a survey with a full factorial design would allow for full examination between all variables, such a survey would require each participant to evaluate 729 choice sets, which is an unrealistic task to perform. Most researchers performing choice experiments thus prefer to use an orthogonal fractional factorial design. This design accommodates the representation of each level of each attribute enough times to document its individual effects, while assuming that the attributes do not interact with each other (Montgomery, 1997). Such designs have been extensively reviewed by Street et al. (2006), and have been used in previous discrete choice research pertaining to inputs for CIMS on subjects ranging from co-generation (Rivers, 2003) to transportation mode choice (Horne, 2003). For each of our surveys, we used an orthogonal fractional factorial design of 18 choice sets, representing the three levels of each of the attributes enough times to document their individual effects on consumer decisions without interactions between attributes. This number of choice sets is well within a participant's ability to provide quality answers according to a study by Hensher et al. (2001), who found that the quality of answers from choice experiments began to deteriorate after the 30th choice due to participants' loss of interest or fatigue. The design matrix for the hybrid vehicle study is presented in Table 2, and that for the hydrogen fuel cell vehicle study is presented in Table 3. We selected the profiles of each choice set manually, based on minimum overlap of attributes in other choice sets, while taking into account the plausible combinations of attributes that a market with the new vehicle technology may experience. In both the HFCV and HEV experimental designs, the capital cost and fuel cost variables specific for the conventional vehicles, and the subsidy, cruising range, refuelling convenience, and warranty variables all had three levels: Low (0), medium (1), and high (2). In the HEV study, the subsidy, cruising range, and warranty attributes for the conventional gasoline vehicle were held constant (C) while the fuel cost variable for the HEV vehicle was linked to the fuel cost of the conventional gasoline vehicle (L) in all choice sets. As for the HFCV study, the subsidy, and warranty attributes for the conventional gasoline vehicle and the fuel cost variable for the HFCV vehicle were all held constant (C) in all choice sets. Street et al. (2006) explored other design strategies that could have yielded higher efficiencies. However we did not pursue these strategies, as they would have produced profile combinations that are not plausible, or would have given rise to too many choice sets for participants to evaluate effectively. Characterizing how consumers' decisions toward HEVs and HFCVs change with different market conditions was a major innovation of our research. In our studies, we focused on the market conditions represented by four market shares of HEVs or HFCVs (0.03%, 5%, 10%, and 20%). We chose these particular market conditions for two reasons. First, policymakers could gain the most from understanding consumer behaviour in the initial stages of a technology's introduction to the market. Second, the market conditions we selected span an appropriate range of market shares that describe the most radical changes in consumer acceptance of a technology. To simplify the language in the survey for participants and to simplify our models, we approximated these market shares using market share ratios, which is the proportion of HEVs or HFCVs to conventional gasoline vehicles. We consider the use of these ratios justified, given the negligible market penetration of other alternative fuel vehicles at the time of our studies.

5 508 ECOLOGICAL ECONOMICS 68 (2008) Table 2 Experimental design matrix for the hybrid electric vehicle experiment Capital Cost Fuel Cost Subsidy Cruising Range Warranty Option 1: conventional vehicle 0 0 C C C 0 1 C C C 0 2 C C C 1 0 C C C 1 1 C C C 1 2 C C C 2 0 C C C 2 1 C C C 2 2 C C C 0 0 C C C 0 1 C C C 0 2 C C C 1 0 C C C 1 1 C C C 1 2 C C C 2 0 C C C 2 1 C C C 2 2 C C C Option 2: hybrid electric vehicle 0 L L L L L L L L L L L L L L L L L L , 1, 2 = Levels, C = Constant, L = Linked to a constant ratio. For each vehicle study, the experimental design accommodated the portrayal of differences in market conditions, which would allow for the independent estimation of discrete choice models for each hypothetical market share. We treated the market share ratio (or market condition) as a blocking variable, separating the survey sample for each study into four market share groups. Each market share group was then assigned a different version of a stated preference survey reflective of their market share ratio. Most components of the four versions of the surveys for each study were designed to be identical, including the attributes and levels in the choice sets. The key difference in the surveys administered to each market share group was the representation of their respective market share conditions. Stated preference surveys generally educate participants about the options presented to them through conceptual descriptions. In our surveys, we facilitated the process of participant education by actively engaging participants in learning about their assigned market conditions. Consumers usually become informed about new products over time, through talking to friends, learning about them in retail outlets, and reading articles in news media and consumer publications. In the survey environment, advanced computer multimedia technology experiences comparable to those in real life can be easily created and delivered to participants in a short period and enable them to gain a good understanding of the technology, both in how the technology works and its effects on society (Peterson et al., 1997). Marketing firms use such methods to inform consumers of new products. An example of these efforts includes the presentation of electric vehicles in virtual showrooms and simulating word-of-mouth communications (Urban et al., 1997, 1996). In our studies, we used a web-based environment to reproduce the experiences associated with the HEV or HFCV technologies by: (1) introduced HEVs and HFCVs to survey participants, (2) controlled for lurking variables variables that consumers may consider, but are not included in the choice Table 3 Experimental Design Matrix for the Hydrogen Fuel Cell Vehicle Experiment Capital cost Fuel cost Subsidy Refueling Convenience Warranty Option 1: conventional vehicle 0 0 C C C 0 1 C C C 0 2 C C C 1 0 C C C 1 1 C C C 1 2 C C C 2 0 C C C 2 1 C C C 2 2 C C C 0 0 C C C 0 1 C C C 0 2 C C C 1 0 C C C 1 1 C C C 1 2 C C C 2 0 C C C 2 1 C C C 2 2 C C C Option 2: hydrogen fuel cell vehicle 0 C C C C C C C C C C C C C C C C C C , 1, 2 = Levels, C = Constant, L = Linked to a constant ratio.

6 509 experiment, and (3) informed participants about their hypothetical market condition. For example, survey participants were able to browse through a brochure (on HEVs or HFCVs, depending on their respective survey) and receive fictional appraisals of HEVs or HFCVs from different people, with variable degrees of technology enthusiasm Data analysis The HEV study collected 916 completed surveys, whereas the HFCV study collected In both cases, survey participants were assigned to one of four market share groups at random, with each group receiving at least 200 completed sets of responses (3600 observations). The regional distribution of the two samples closely reflected that of Canada in 2002 in terms of population, income, age, and household size. With the exception of a slightly higher representation of women in the HFCV study than the HEV study, all other demographic characteristics and trends in vehicle ownership among market share groups in both studies match strongly, limiting the introduction of significant bias in our results. Both vehicle studies required the estimation of four discrete choice models from the data obtained through the choice experiments, one for each hypothetical market share. Our studies drew separate samples from the population of interest. To avoid bias, individuals who participated in the HEV study did not participate in the HFCV study. Analysis of survey results was also an independent process, resulting in separate multinomial logit (MNL) models for the two studies. The distinct advantage of running parallel studies is that we are able to compare the results of respective discrete choice models to draw inferences on differences or similarities in consumer decision-making with respect to an evolutionary type of technology (the HEV) and a disruptive technology (the HFCV). Mathematically, discrete choice models define the probability (P in ) that a consumer (n) will choose an alternative (i) froma set of alternatives (J n ) as dependent on the observed characteristics (z) of alternative i compared to all other alternatives, and on the observed characteristics of the consumer (S n ) as in Eq. (4) (Train, 1986): P in ¼ f z in ; z jn for all j in J n and jpi; s n ; b ð4þ Here j represents possible choices in the set of alternatives J. The observed characteristics are defined by a vector of parameters (β), representing the importance consumers place on each characteristic in the alternatives. For example, in a discrete choice model of vehicles, β may represent the relative importance consumers place on a vehicle's price, warranty period, and body type. We can relate this vector of parameters to economic theory by representing the relative importance of characteristics as an index of consumer value in units of utility. We can estimate each alternative's total utility by using a function that aggregates the utility consumers derive from each characteristic. In our study, we assumed that utility pertaining to Canadians' vehicle choices is linear-in-parameters, which was shown to be a reasonable approach by another recent vehicle study (Horne et al., 2005). This utility function is illustrated in Eq. (5): U j ¼ b j þ XK k¼1 b k x jk þ e j Here U j is the utility of technology j, β j is the alternative specific constant, β k is a vector of weighing coefficients, representing the importance that consumers place on attribute k, x jk is a vector of the k attributes of technology j, and e j is the unobservable error term. The attributes are assumed to be continuous variables in the utility function, since costs and warranty periods presented to a consumer are rarely ordinal in a purchase situation. Based on the attributes included in the HEV and HFCV surveys, we can represent Eq. (5) as Eq. (6) for the HEV study, and as Eq. (7) for the HFCV study. U HEV ¼ b CC CC þ b FC FC þ b CR CR þ b SUB SUB þ b W W þ b HCONV þ e U HFCV ¼ b CC CC þ b FC FC þ b RC RC þ b SUB SUB þ b W W þ b HASN þ e Here U HEV is the utility calculated from the HEV study, and U HFCV is the utility calculated from the HFCV study; CC is capital cost (sale price of the vehicle), FC is weekly fuel cost, CR is the cruising range, RC is refuelling convenience, SUB is government subsidy, W is warranty coverage, e is the unobservable error term, β HCONV is the constant specific to conventional gasoline vehicles in the HEV study; and β HASC is the constant specific to hydrogen fuel cell vehicles in the HFCV study. In our research, we assumed that the unobservable error term for each alternative was distributed independently and identically in accordance with the extreme value distribution. While it is very difficult and unrealistic to guarantee that this assumption has been satisfied, as with all models of this type, we have carefully selected attributes that likely dominate a consumer's decision in purchasing vehicles in previous studies (Horne et al., 2005; Ewing and Sarigöllü, 2000; Brownstone et al., 2000; Brownstone and Train, 1999; Greene, 1997; Bunch et al., 1993). The extreme value distribution approximates many situations well in social and natural sciences, including consumer behaviour (Horne, 2003; Train 1986, 2003). These models express the probability that the decision maker will choose alternative i over all alternatives J as in Eq. (8), where V in and V jn denote the observable portion of utility for these technologies. Discrete choice models using this distribution are called multinomial logit (MNL) models. P in ¼ ev in P J e V jn j¼1 ð5þ ð6þ ð7þ ; For all i in J n ð8þ Using maximum likelihood estimation, we then calculated the β parameters for each attribute in the HEV study using Eq. (6) and the HFCV study using Eq. (7), which, for each study, would allow us to balance Eq. (8) for a given set of consumer choice data. This technique finds the value of β parameters that are most likely, given the actual discrete choices made by survey participants. In this research, we estimated MNL models from data collected in both studies, for each market share group. Although the MNL

7 510 ECOLOGICAL ECONOMICS 68 (2008) models can provide useful information, translating their outputs into parameter values for use in CIMS simulations allows for a more comprehensive and realistic simulation of policy options. The MNL model used in this research is one of several similar models that can estimate consumer behaviour parameters. We have chosen the MNL model for its simplicity and its ability to seek approximate results quickly in transportation applications. The trade-off in using this simple model specification is in its inability to take into account random variations in consumer tastes, such that the MNL model assumes each estimated coefficient to be identical for every consumer (IID assumption). More complex models, such as latent class and mixed MNL models would provide results that would overcome the IID assumption (Train, 2003). Given that the purpose of the MNL model outcomes in this study is to estimate approximate and general behavioural parameters for CIMS at an aggregate level, and that CIMS does not yet account for variations in consumer tastes, we accepted the limitations of the MNL model. We recommend that future research building on this study with the aim to further refine the behavioral parameters in CIMS should consider using one of these more complex models to provide a more precise characterization of consumer preferences Integrating the results into CIMS Translating the findings on consumer behaviour from the MNL models to parameters in CIMS required three steps for each technology. First, we estimated parameters in CIMS' intangible cost function from the utilities for non-monetary attributes in the MNL models. Parameters A and k of the intangible cost function in Eq. (3) are estimated exogenously to CIMS simulations, whereas NMS is an endogenous variable. To capture dynamics in intangible costs under different market share conditions, we estimated A and k from each MNL model based on technology market shares. Using odds ratios, we compared the non-cost parameters such as cruising range and refuelling convenience to annual cost parameters, so that these intangible (non-monetary) variables could be compared and discussed in dollar terms: the example for a HEV and a conventional gasoline vehicle (Eq. (6)), we estimated the private discount rate as follows: r ¼ b CC b FC T52 1 ð1 þ r Þ n ð10þ Where β CC is the parameter for capital cost, taken directly from the MNL models; andβ FC is the coefficient for weekly fuel costs from the MNL models. β FC, the weekly fuel cost coefficient, is scaled by a factor of 52 to provide an annualized fuel cost coefficient. Variable n is the lifespan of a vehicle, which here we estimated to be 16 years. The third CIMS behavioural parameter (v), represents the degree of market heterogeneity, and approximates the scale of the MNL model or the relative magnitude of the error term to MNL parameter values. We can only estimate v if the lifecycle costs (LCC) of all technology alternatives are already known. That is, all of the components of the LCC, whether they are static or are represented as functions, must be defined. This includes the capital cost, discount rate, maintenance cost, energy cost, and intangible costs of each technology in the same competition node. A major assumption in estimating the v parameter is that both CIMS and the MNL model yield the same probabilities of consumers choosing one technology over another, or, that these two models will forecast the same market shares of technologies. Following this assumption, we equate the MNL model to a node of technology competition in CIMS, as in Eq. (11). All variables in both models except for the v parameter in CIMS are known or are solved from data inferred from the discrete choice research, as explained above. Thus, we can estimate v, the remaining parameter by using linear approximation. (11) m i ¼ b i b CC Ti ð9þ 4. Results and discussion 4.1. Discrete choice models In Eq. (9), m i is the monetized value of attribute i, β i is the estimated parameter value for attribute i in a given utility function, and β cc is the estimated parameter value for capital cost in the utility function. Note that the parameters β i and β CC are weighing coefficients, relating their respective attributes to utility, and thus, β i is in units of utility per i, and β CC is in units of utility per dollar. We approximated the A and k parameters of CIMS' intangible cost function by fitting the total intangible costs of each market condition to the intangible costs curve defined by Eq. (3). Second, we estimated the private discount rate, which represents consumers' time preferences with respect to financial costs. The private discount rate is determined by comparing the capital cost parameter from the utility functions in the MNL models to annual cost parameters, reflecting ongoing costs over the lifespan of the vehicle technology (Horne et al., 2005). Using We built multinomial logit (MNL) models from choice data collected in both studies, for each market share group, according to Eqs. (6) and (7). We then used the software package LIMDEP 9.0 to find the maximum likelihood estimators (MLE) for each parameter in the MNL models. Tables 4 and 5 show the MNL modeling results from the HEV and HFCV studies for all market share groups, including the attribute parameters (β) as well as an indication of their statistical significance using T-statistics. As previously mentioned, the four market share groups correspond to different degrees of hypothetical market penetration of HEVs relative to conventional gasoline vehicles (Table 4), and HFCVs relative to conventional gasoline vehicles (Table 5). Market share group 1 (MS1) assumes a market penetration of the new vehicle technology of 0.03%; MS2 assumes a penetration of 5%, MS3 of 10%, and MS4 of 20%.

8 511 Table 4 MNL model parameters from HEV study Attribute Market share 1 (MS1) model (0.03 percent) MS2 model (5 percent) MS3 model (10 percent) MS4 model (20 percent) ß t-ratio ß t-ratio ß t-ratio ß t-ratio Capital cost 1.95E E E E Fuel cost 4.12E E E E Government subsidy 2.90E E E E Cruising range 7.91E E E E Warranty coverage 1.66E E E E ASC Conventional gasoline vehicle 2.02E E E E Observations Log likelihood full model (F) Log likelihood constants only (ASC) Log likelihood no coefficients (0) (L(F) L(0)) (L(F) L(ASC)) Likelihood ratio index: 1 L(F) /L(ASC) Likelihood Ratio Index: 1 L(F) /L(0) All coefficients are significant at the 95% confidence level. The full model is significantly different than a null model at 99.9% confidence level ( ). The full model is significantly different than a model with only the ASC at 99.9% confidence level ( ). We evaluated the HEV and HFCV models' explanatory powers statistically by testing whether they are better than models with all coefficients equal to zero (null model), or models with only the alternative specific constant (restricted model) by comparing the log-likelihoods of each of these models. The log-likelihood of a model is a negative variable that approaches zero with model validity, such that a discrete choice model capable of consistently predicting the actual choice made would have a log likelihood of zero. We tested whether there are significant differences between (1) the full model and the restricted model, and (2) the full model and the null model by making use of the fact that twice the difference between the full and restricted models is chi-squared distributed, with the number of degrees of freedom equal to the number of explanatory variables in the full model minus the number in the restricted model and in the null model, respectively. Using these tests, we found that the full model for both the HEV and HFCV studies are better fits than either the restricted or null models, since twice the difference in their log likelihoods exceeded the critical value of the chi-squared Table 5 MNL model parameters from HFCV study Attribute Market share 1 (MS1) model (0.03 percent) MS2 model (5 percent) MS3 model (10 percent) MS4 model (20 percent) ß t-ratio ß t-ratio ß t-ratio ß t-ratio Capital cost 1.47E E E E Fuel cost 1.96E E E E Government subsidy 2.87E E E E Refuelling convenience 9.04E E E E Warranty coverage 1.37E E E E ASC Hydrogen fuel cell vehicle 7.32E E E E Observations Log likelihood full model (F) Log likelihood constants only (ASC) Log likelihood no coefficients (0) (L(F) L(0)), Chi Square (L(F) L(ASC)), Chi Square Likelihood Ratio Index: 1 L(F)/ L(ASC) Likelihood Ratio Index: 1 L(F)/ L(0) All coefficients are significant at the 95% confidence level with the exception of fuel cost for MS2 ( ). The full model is significantly different than a null model at 99.9% confidence level ( ). The full model is significantly different than a model with only the ASC at 99.9% confidence level ( ).

9 512 ECOLOGICAL ECONOMICS 68 (2008) Fig. 1 HEV significant coefficients at the 95% interval between the lowest and highest market shares (Error bars represent 95% confidence interval). distribution with the appropriate number of degrees of freedom, at the 99.9% confidence level (Tables 4 and 5). Next, we looked at the predictive capacity of the qualitative models through the log likelihood ratio index. This index is commonly used to assess the goodness of fit of the models to their respective data (Train, 2003). The index ranges from 0 to 1; the value of 0 indicates that the estimated model has no predictive power, whereas a value of 1 indicates that the model can predict all consumer decisions perfectly. We found that in both the HEV and HFCV studies, this index had a value between 0.15 and 0.30 when the full models were compared with the restricted and null models, suggesting that the full models have much better predictive capacity. These values are also within the same range as the log likelihood ratio index values reported in other similar qualitative transportation models (Ewing and Sarigöllü, 2000) Consumer preferences for hybrid gasoline-electric vehicles Analysis of results from the HEV study support our hypothesis that changes in market share affect consumer behaviour and increase consumers' propensity to choose HEVs. Specifically, three attribute parameters from the HEV study exhibit trends as HEVs gain market share, between the lowest and highest market shares tested using the standard t-test at 95% confidence intervals (Fig. 1): the standardized value of the alternative specific constant (ASC), the perceived undesirability of higher fuel prices, and the perceived benefit of longer cruising range. The trends for these attributes provide strong evidence that consumer preferences for HEV are dynamic, and are a function of HEV market share. In other words, the HEV data are evidence of the neighbor effect. The ASC captures the additional perceived benefits or costs associated with conventional gasoline vehicles over HEVs due to attributes not considered in the choice experiment. We found that with an increasing HEV market share from 0.03% to 20%, the value of attributes explained by the ASC declines (Table 6). This trend suggests that consumers view conventional gasoline vehicles as becoming less desirable with increases in HEV adoption, all else being equal. Factors such as performance, reliability, and safety of HEVs could partly explain this trend. These are factors that consumers may take into account in their vehicle purchase decisions (Horne et al., 2005; Ewing and Sarigöllü, 2000), which we did not explicitly model in this study. Aside from implications related to the ASC, Table 6 also indicates that as the market share of HEVs increases from 0.03% to 20%, higher fuel prices becomes less of a concern, and Table 6 Value of an additional unit of an attribute in terms of capital cost (HEV) Attribute Equivalent perceived change in terms of capital cost for each market share of HEV MS1 (0.03 percent) MS4 (20 percent) ASC (value of purchasing a conventional gasoline vehicle instead of a HEV) One dollar increase in monthly fuel cost One more day of typical driving on the same amount of fuel $1000 $ 3000 $210 $100 $400 $160 Fig. 2 Net difference in the effects of non-monetary attributes for a Honda Civic hybrid over the conventional Honda Civic in dollar equivalents of utility ( consumer value index ).

10 513 Table 7 Value of an additional unit of an attribute in terms of capital cost (HFCV) Attribute Equivalent change in terms of capital cost for each market share of HFCV ASC (value of purchasing a HFCV instead of a conventional gasoline vehicle) One dollar increase in monthly fuel cost One more station with the proper system for refuelling HFCVs One additional year of warranty coverage One dollar increase in government subsidy MS1 (0.03 percent) MS4 (20 percent) $50,000 $52,000 $133 $90 $61,500 $67,000 $932 $1182 $1.95 $2.17 longer cruising range becomes less of an attractive feature to the general commuter population, as represented in our survey sample. As more consumers choose to adopt HEVs in their next car purchase under higher HEV market shares, this same group of consumers becomes less concerned about fuel costs because HEVs are fuel efficient. An examination of the choices made by the survey participants concludes that the instances in which survey respondents chose HEVs increased from 49% (MS1) to 56% (MS4). Combining all of the dynamic effects reveals two additional observations as the market share of HEVs increases: the importance consumers place on the non-monetary attributes generally decline, and the net consumer value of nonmonetary attributes for the HEV is always higher than that of the conventional gasoline vehicle. These two observations indicate that consumers' preferences for HEVs become stronger as more people drive HEVs, given equal monetary costs of the two vehicle technologies. Fig. 2 presents this trend, with the consumer value index representing consumer values perceived as a change in the vehicle's total cost Consumer preferences for hydrogen fuel cell vehicles In contrast to the results of the HEV study, the MNL models describing consumer preferences for HFCVs do not support the neighbor effect. The results indicate that the HFCV choice is highly dominated by the alternative specific constant (ASC) and stations with the proper fuel. That is, the magnitude of the ASC and refueling convenience and their statistically significant contribution to utility in the models of the HFCV study are high. Taken in the context of our experiment, this suggests that HFCVs can achieve significant market shares just by virtue of being HFCVs, or by constructing a few more refuelling stations, all else being equal (Table 7). The model predictions for HFCVs include much uncertainty, and it is clear that the survey respondents were attracted to this vehicle type for reasons that cannot be fully explained by our basic model specifications. Several factors could account for HFCV's dominant appeal, three of which could be: (1) respondents overstated their preference for HFCVs because of a perceived social good aspect; (2) the omission of attributes in the choice experiment that are important to decision-making; and (3) respondents' reaction to disruptive technologies. Researchers assessing the marketability of environmentally friendly technologies or products often find a discrepancy between respondents' stated attitudes towards an environmental or social outcome they perceive as desirable (such as improving air quality) and their revealed behaviour (such as buying fuel inefficient vehicles and neglecting alternative modes of transport) (Roberts and Bacon, 1997). Indeed, people's preferences tend to vary according to whether, at the moment of observation, they perceive their decision to affect their personal utility or society at large (Sagoff, 1988). During the consumer education portion of the survey, respondents received information on the low emissions of HFCVs compared to gasoline vehicles. In this way, the information provided to respondents on HFCVs might have placed gasoline vehicles at a competitive disadvantage given that there was no mention of advances in conventional gasoline vehicles that could result in substantial fuel efficiencies. Thus, the information might have had a strong effect on respondents' frames of mind during the survey, putting the public social / environmental good in the forefront, though only temporarily. A second possible explanation has to do with attributes that we did not include in the choice experiment. The intent of the education part of the survey was to give respondents the sense that HFCVs could compete with gasoline vehicles on various aspects, but due to research constraints we were unable to explore other aspects such as the diversity of vehicle makes and models available on the market. Some previous studies have concluded that a limited selection of body types relative to conventional gasoline vehicles is a key factor preventing alternative fuel / technology vehicles from achieving significant market penetration (Leiby and Rubin, 2003). Excluding this constraint might have artificially increased the appeal of HFCVs, specifically for respondents that are partial to certain vehicle body types or brands. However, both of these shortcomings in research methods also apply to the hybrid gasoline-electric vehicle (HEV) study, the results of which do not yield the same dominance of the ASC. Comparisons such as these lead us to speculate that the dominant appeal of HFCVs relative to gasoline vehicles relates to the radically new features of HFCVs. People might be attracted to the radical newness of this vehicle type, such as the revolutionary drive train and fuel. Perhaps people attach a perceived status to owning this vehicle technology. At the same time, people's lack of familiarity with HFCVs could have led them to assume that the potential negative attributes were negligible, Table 8 Baseline values for vehicle attributes (Honda 2004) Honda Civic Honda Civic hybrid Capital cost ($) $17,100 $29,510 Fuel cost ($/week) Government subsidy ($) Cruising range (days/tank) Warranty (years) 5 5 Alternative specific constant for gasoline vehicles 1 0

11 514 ECOLOGICAL ECONOMICS 68 (2008) Table 9 Private discount rates estimated for HEV market share groups Market shares of HEVs 0.03% 5% 10% 20% 21% 28% 35% 49% whereas in the HEV study respondents might have had some information on the negative attributes associated with HEVs. The dominance of the ASC for HFCVs could reflect the potential for dramatic switches to this vehicle technology, once it has attained a given level of development and acceptability. Christensen described the commercialization of some disruptive technologies in the computing industry, and explained how established companies can fail to predict the mainstream appeal of disruptive technologies, since these technologies tend to satisfy only niche market segments at the outset (Christensen, 1997; Bower and Christensen, 1995). By definition, disruptive technologies present a set of attributes that existing customers might not value initially. However, improvements in valued and new attributes can rapidly match and outpace customers' demands, making it possible for the new technology to penetrate the primary market. The potential for this phenomenon to occur with (hydrogen) fuel cell vehicles might explain why established vehicle manufacturers, such as Daihatsu, Daimler-Chrysler, Fiat, Ford, GM, Honda, Hyundai, Mazda, Mitsubishi, Nissan, Peugeot/Citroën, Renault, Toyota, and Volkswagen, are exploring prototypes of technology alternatives that focus on fuel cells (US Department of Energy, 2004). Due to the uncertainty in the results of the HFCV study, the remainder of the paper will focus on using the findings on consumer preferences for hybrid electric-gasoline vehicles to improve the ability of the energy-economy model CIMS to simulate long-term technological change New parameters for the CIMS model We used the results of our discrete choice research to update parameters in CIMS so that the simulations of technological change include preference dynamics. This innovation enables more realistic simulations by enhancing the feedback loops between consumer preferences, financial costs, and market shares, allowing new technologies to garner market shares over the long-term. To illustrate the application of our research contribution to CIMS simulations, we used the input assumptions for gasoline and HEV shown in Table 8. Using Eq. (10) and assuming that, similar to conventional vehicles, HEVs have a technological lifespan of 16 years, the results show that the private discount rate is dynamic and is affected by changes in market conditions (Table 9), increasing from 21% to 49%. This indicates that as more HEVs are present on the road, consumers become less concerned about the fuel cost of HEVs, but relatively more concerned about the capital costs of the car. Other studies have also found consumers to be most sensitive to capital costs amongst other attributes when purchasing a vehicle, leading to high discount rates. Our estimates for the private discount rate are consistent with those of other vehicle choice studies: Horne et al. (2005) reported a discount rate of 22.6%, and Train (1985) and Ewing and Sarigöllü (2000) also reported discount rates in the same range. CIMS cannot accommodate the dynamic nature of the private discount rate revealed by our HEV research because r is constant throughout all periods of a simulation run. To demonstrate CIMS' new modeling capabilities we chose the discount rate of 21% for policy simulations, which reflects the current market condition of hybrid gasoline-electric vehicles. According to the method outlined in the previous section, the non-cost attributes in the MNL models under each market share condition are translated into the declining intangible cost function in CIMS. Table 10 outlines the parameters estimated for this function for hybrid electric-gasoline vehicles Implications for policy simulations The incorporation of dynamic consumer preferences in CIMS can affect the outcomes of policy simulations. We illustrate this with two simulations of a reference case using the characteristics of a Honda Civic and the Honda Civic hybrid. One simulation used the behavioural parameters developed in this research, and the other did not. With the exception of behavioural parameters, both simulations are identical in their starting assumptions, including an identical node structure and economic conditions. Fig. 3 below shows that when dynamic behavioural parameters are included, CIMS forecasts the national total market shares of hybrid gasolineelectric vehicles to improve over time. Without the dynamics in behavioural parameters, the market share forecast of hybrid vehicles is stagnant. Thus, the static portrayal of consumer preferences in long-run simulations is likely to Table 10 Parameters for CIMS dynamic intangible cost function (HEV) Case Parameter Type Io A k HEV Baseline $ Io = initial intangible cost; A and k respectively define the shape of the curve and the rate of change. Fig. 3 CIMS reference projections of the national total market share of HEVs.

12 515 underestimate the penetration of new technologies, such as hybrid gasoline-electric vehicles. 5. Conclusions This research has demonstrated that consumers' preferences in choosing between conventional and new technologies can change with market conditions. The importance that consumers place on certain attributes of a new technology also changes as the new technology gains market share. These findings support the importance in the ability of energy-economic models to capture such consumer behaviour, thereby increasing the realism of simulation outputs and their usefulness in the assessment of long-run policy alternatives. These findings also provide a potential rationale for government support for new technologies which offer positive environmental impacts, such as reductions in greenhouse gas emissions. Additionally, our research suggests that dynamics in consumer preferences also depend on the type of new technology, such that preference patterns for evolutionary technologies are unlikely to apply to disruptive technologies. The HEV results produce consistent trends as market conditions for HEVs improve, while the results for HFCVs are much less consistent. This inconsistency may be due to our challenge in capturing attributes that are specific to choices between conventional vehicles and HFCVs. The vehicle attributes included in our research were based on consumers' preferences for conventional vehicles. While this may be acceptable when considering the penetration of HEVs, consumers may be concerned about vastly different attributes when HFCVs become a viable choice in their vehicle purchase decision. This is an important area for future research. Acknowledgments We would like to thank Jacqueline Sharp, Nic Rivers, John Nyboer, Alison Laurin, Rose Murphy, Kevin Tu, and Chris Bataille for your helpful input, guidance and advice in this research. We would also like to thank the Office of Energy Efficiency at Natural Resource Canada, the National Science and Engineering Research Council, Simon Fraser University School of Resource and Environmental Management, the Canadian Institute of Energy, and Environment Canada for the funding, which made this research possible. REFERENCES Brownstone, D., Bunch, D.S., Train, K., Joint mixed logit models of stated and revealed preferences for alternative-fuel vehicles. Transportation Research B34, Brownstone, D., Train, K., Forecasting new product penetration with flexible substitution patterns. Journal of Econometrics 89, Bunch, D., Bradley, M., Golob, T., Kitamura, R., Occhuizzo, G., Demand for clean-fuel vehicles in California: a discrete-choice stated preference pilot project. Transportation Research 27A (3), Bower, J.L., Christensen, C.M., Disruptive technologies: catching the wave. Harvard Business Review Jan/Feb. Christensen, C., Innovator's dilemma: When new technologies cause great firms to fail. Harvard Business School Press, Boston, Mass. DeShazo, J.R., Fermo, G., Designing choice sets for stated preference methods: the effects of complexity on choice consistency. Journal of Environmental Economics and Management 44, Ewing, G., Sarigöllü, E.S., Assessing consumer preferences for clean-fuel vehicles: a discrete choice experiment. Journal of Public Policy & Marketing 19 (1), Fujii, S., Garling, T.G., Application of attitude theory for improved predictive accuracy of stated preference methods in travel demand analysis. Transportation Research Part A 37, Greene, D.L., Survey evidence on the importance of fuel availability to choice of alternative fuel vehicles. Energy Studies Review 8 (3), Hensher, D., Stopher, P., Louviere, J.J., An exploratory analysis of the effect of the number of choice sets in designed choice experiments: an airline choice application. Journal of Air Transport Management 7 (6), Hensher, D., Louviere, J., Swait, J., Combining sources of preference data. Journal of Econometrics 89, Horne, M., Jaccard, M., Tiedemann, K., Improving behavioral realism in hybrid energy-economy models using discrete choice studies of personal transportation decisions. Energy Economics 27 (1), Horne, M Incorporating preferences for personal urban transportation technologies into a hybrid energy-economy model. Master's Thesis. School of Resource and Environmental Management. Burnaby, BC, Simon Fraser University: 190pp. Jaccard, M., Dennis, M., Estimating home energy decision parameters for a hybrid energy economy policy model. Environmental Modeling and Assessment 11 (6), Jaccard, M., Murphy, R., Rivers, N., Energy-environment policy modelling of endogenous technological change with personal vehicles: combining top-down and bottom-up methods. Ecological Economics 51, Jaccard, M., Nyboer, J., Bataille, C., Sadownik, B., Modeling the cost of climate policy: Distinguishing between alternative cost definitions and long run cost dynamics. The Energy Journal 24 (1), Jackson, J., Disruptive technologies. Washington Technology 17 (20). Leiby, P., Rubin, J., Understanding the transition to new fuels and vehicles: Lessons learned from analysis and experience of alternative fuel and hybrid vehicles. 19 pp. Mackay, M.M., Metcalfe, M., Multiple methods forecasts for discontinuous innovations. Technological Forecasting and Social Change 69, Murphy, R., Analysis of measures for reducing transportation sector greenhouse gas emissions in Canada, Masters thesis. School of Resource and Environmental Management. Burnaby, British Columbia, Simon Fraser University. Montgomery, D.C., Design and analysis of experiments, Fourth Edition. Wiley, New York Nyboer, J., Simulating evolution of technology: An aid to energy policy analysis, Phd thesis. School or Resource and Environmental Management. Burnaby, British Columbia, Simon Fraser University. Peterson, R., Balasubramanian, S., Bronnenberg, B.J., Exploring the implications of the internet for consumer marketing. Academy of Marketing Science Journal 25 (4), 329. Rivers, N., Jaccard, M., Combining top-down and bottom-up approaches to energy-economy modeling using discrete choice methods. The Energy Journal 26 (1),

13 516 ECOLOGICAL ECONOMICS 68 (2008) Rivers, N Behavioural realism in a technology explicit energy-economy model: The adoption of industrial cogeneration in Canada. Master's Thesis. School of Resource and Environmental Management. Burnaby, BC, Simon Fraser University: 146. Roberts, J.A., Bacon, D.R., Exploring the subtle relationships between environmental concern and ecologically conscious consumer behavior. Journal of Business Research 40 (1), Sagoff, M., The Economy of the Earth. Cambridge University Press, Cambridge. Street, D., Burgess, L., Louviere, J.J., Quick and easy choice sets: Constructing optimal and nearlyoptimal statedchoice experiments. International Journal of Research in Marketing 22, Train, K., Discount rates in consumers' energy-related decisions: A review of the literature. Energy 10, Train, K., Qualitative choice analysis Theory, econometrics, and an application to automobile demand. The MIT Press, Cambridge, MA. Train, K., Discrete choice methods with simulation. Cambridge University Press Urban, G.L., Hauser, J.R., Qualls, W.J., Weinberg, B.D., Bohlmann, J.D., Chicos, R.A., Information acceleration: Validation and lessons from the field. Journal of Marketing Research 34, Urban, G.L., Weinberg, B.D., Hauser, J.R., Premarket forecasting of really-new products. Journal of Marketing 60, US Department of Energy, Fuel cell vehicle world survey Breakthrough Technologies Institute, Washington DC. February 2004.