Research in Transportation Economics 50 (2015) 17e28

Contents lists available at ScienceDirect

Research in Transportation Economics

journal homepage: www.elsevier.com/locate/retrec

The choice of Bass model coefficients to forecast diffusion for

innovative products: An empirical investigation for new automotive
technologies
^ me Massiani a, *, Andreas Gohs b

ro
Je
a  Ca' Foscari di Venezia, Dipartimento di economía, Cannaregio 873, 30121 Venezia, Italy
Universita
b €t Kassel, Germany
Universita

a r t i c l e i n f o a b s t r a c t

Article history: Bass diffusion models are one of the competing paradigms to forecast the diffusion of innovative
Available online 10 July 2015 products or technologies. This approach posits that diffusion patterns can be modeled through two
mechanisms: Innovators adopt the new product and imitators purchase the new product when getting in
JEL: contact with existing users. Crucial for the implementation of the method are the values assigned to the
Q55 two parameters, usually referred to as p and q, which mathematically describe innovation and imitation
O33
mechanisms. The present paper is based on the findings of a research project about policy measures to
Keywords: promote the diffusion of Electric Vehicles in Germany. It investigates how practitioners could choose
Bass diffusion model adequate values for the Bass model parameters to forecast new automotive technologies diffusion with a
Innovation focus on Electric Vehicles. It considers parameters provided by the literature as well as ad hoc parameter
Electric vehicles estimations based on real market data for Germany. Our investigation suggests that researchers may be
in trouble in electing adequate parameter values since the different eligible parameter values exhibit
dramatic variations. Literature values appear discussible and widely variable while ad hoc estimates
appear poorly conclusive. A serious problem is that ad-hoc estimates of the Bass p value are highly
sensitive to the assumed market potential M. So for plausible values of M, p varies on a high scale. Unless
more consolidation takes place in this area, or more confidence can be placed on ad hoc estimates, these
findings issue a warning for the users of such approaches and on the policy recommendations that would
derive from their use.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction the diffusion model a  la Bass still lack a form of consolidation in

that, to our knowledge, no effort has been made, at least in our
With the increasing popularity of Electric Vehicles (EVs) field, to compare the underlying parameters in a systematic
among policy makers and the general public, economists have manner and to assess how analysts can select parameter values to
been called to design convincing methods for forecasting the conjecture market diffusion scenarios. These parameters known
development of this market. Among the various competing par- as p and q, or coefficients for innovation and imitation represent
adigms, Bass diffusion theory received attention from scientists the intrinsic driving forces, together with potential M, of the
and has been applied in research projects as is reported in a diffusion pattern of the innovative technology, and thus deserve
number of scientific communications. Yet approaches based on careful consideration.
This paper aims at filling in this gap, as it provides a systematic
examination of the p and q parameters proposed in the literature
as well as parameter estimates based on real market data. Based
on extensive research and state of the art analysis performed
* Corresponding author.
during a research project about policy measures to promote the
E-mail address: [email protected] (J. Massiani). diffusion of Electric Vehicles in Germany, our initial intention was

http://dx.doi.org/10.1016/j.retrec.2015.06.003
0739-8859/© 2015 Elsevier Ltd. All rights reserved.
18 J. Massiani, A. Gohs / Research in Transportation Economics 50 (2015) 17e28

to provide guidelines for practitioners in their choice of parameter

values for EV diffusion forecast. However, our findings raise
skepticism on the possibility to convincingly model EV market
diffusion through the Bass model, considering the dramatic
discrepancy between reported estimates and the inconclusiveness
of ad hoc estimates.
The paper proceeds as follows: In the next section, the Bass
model for innovation diffusion is presented. The third section in-
vestigates the estimates provided by literature about the value of p
and q parameters for innovative automotive technologies. The
fourth section presents the results of our own estimations based on
real market data collected for the German market. The conclusion
puts into perspective these different results and underlines the
difficulty which researchers and practitioners may encounter when
electing the Bass parameter values.

Fig. 1. A typical textbook illustration of diffusion patterns in the Bass model.

2. The Bass model of diffusion

Electric Vehicles (EVs), Hybrid Electric Vehicles (HEVs), Biofuel

Vehicles and, to some extent, Liquid Petroleum Gas (LPG) and
Compressed Natural Gas (CNG) propelled vehicles can be consid- f ðtÞ ¼ ðp þ qFðtÞÞð1  FðtÞÞ (2)
ered as new propulsion technologies which could potentially sub-
stitute established Internal Combustion Engine technologies (i.e. With
Gasoline and Diesel powered vehicles). The adoption of new
technologies has often been found to follow an S-shaped curve. In f(t), density function of sales at time t,
this setting, cumulated purchases of innovative products can be F(t), cumulative fraction of the potential that is achieved at
characterised by three different growth phases: a slow take-up time t.
phase, followed by a phase of more rapid growth as the technol-
ogy becomes widespread and, finally, slowing growth when the Integrating over time, the total fraction of the potential that is
‘not so new’ technology approaches saturation. Diffusion theories adopted at time t is:
try to explain or model the actual shape of diffusion curves e i.e. the
1  e eðpþqÞt
speed and the shape of cumulative adoption of an innovation FðtÞ ¼ (3)
among a set of prospective buyers. A widely used approach in 1 þ pqeðpþqÞt
marketing to explain the diffusion of innovations is the so-called
These premises set, it is all apparent that the diffusion pattern
Bass model (Bass, 1969).
will be highly dependent of the following elements:

2.1. Bass model in synthesis
The estimate of the potential market,

The values used for p and q.
The Bass model assumes that potential buyers of an innovation
can be divided into two groups: Armored with this apparatus, the Bass model has become a
leading paradigm for the forecast of innovative technology diffu-

Innovators: People who buy the product first and are influenced sion. In the subsequent section, we analyze more in detail the
only by ‘external communication’ e.g. mass media or reasons for such a success (Fig. 1).
advertisement.

Imitators: Individuals who, in contrast, buy if others have
already bought the product since they are influenced by word of 2.2. The Bass model: a long lasting success
mouth or so-called ‘internal communication’ (Mahajan, Muller,
& Bass, 1990). The Bass model is routinely used by analysts to explore the
market potential for innovative propulsion technologies. The Bass
Following this narrative, the number of first time purchases nt at model is often applied in studies where diffusion plays a decisive or
time t can be expressed as follows: critical role. For example, diffusion assumptions are key drivers in
energy studies like that of (Lee, Park, Kim, & Lee, 2013), but are
dNt 1 considered with limited attention in this field of research, because
nt ¼ ¼ pðM  Nt Þ þ q Nt ðM  Nt Þ (1)
dt M the main focus is often on energy generation and distribution
issues.
With: The success of the Bass model relates to different factors. First,
the model is authoritative, especially in management sciences. Bass
nt: product purchases in period t seminal paper was selected in 2004 as one of the ten most
Nt: cumulative product purchases until beginning of period t frequently cited papers in the 50-year history of Management Sci-
M: cumulative market potential on the whole product's life cycle ence.1 The rhetoric of Bass approach also insists on the initial
p: coefficient of innovation q: coefficient of imitation.

Alternatively, the model is sometimes written as (this may turn 1

Management Science 50 Number 12 Supplement, Dec 2004 ISSN 0025-1909 p.
out useful looking at some authors who use this notation): 1833e1840.
 J. Massiani, A. Gohs / Research in Transportation Economics 50 (2015) 17e28 19

success of the model: The diffusion patterns predicted by Bass in his of new automotive technologies, it appears useful for us, to
original 1969 article have shown to adhere convincingly with analyze available information about the adequate values for p and
subsequent real market outcomes for the set of technologies q considering two sources: A literature review, and an estimation
considered by the author. Massiani (2012) also analyses other on a data set of new automotive technologies, essentially in
reasons of success of the Bass paradigm in comparison with other Germany.
competing approaches to model the diffusion of Alternative Pro-
pulsion Technologies: Bass diffusion patterns have a good face
value as they draw a smooth diffusion curve, which is consistent
with the common perception of a slow uptake and a saturation 3. Literature review: widely variable values are available for
when demand becomes higher. Additionally, those diffusion pat- Bass parameters
terns are intrinsically made consistent with the time series
observed up to the present time. Additionally, Bass models are data This section reviews the values of p and q parameters found in
parsimonious: A time series of the seminal diffusion of a given literature sources. Several types of sources can be distinguished
technology is usually the only requirement to estimate Bass pa- based on whether they refer to other (partly authoritative) sources
rameters. Also an analyst can rely on preexisting estimates of the or whether they produce an own estimate.
set of p, q and m parameters. Compared to Stated Preference (or
Conjoint Analysis) surveys, that usually require collections of spe-
cific data (except in the proposed Synthetic Utility Function advo-
3.1. Reference to authoritative sources
cated in Massiani, 2012) and more elaborated computations, and to
Total Cost of Ownership that necessitates a large set of engineering
First, part of the literature provides intervals rather than
and accounting data, the Bass approach can constitute an attractive
point values for p and q coefficients. In a review of the Bass
alternative. Bass, Krishnan, and Jain (1994) also show that the Bass
literature, Chandrasekaran and Tellis (2007) refer to p values
model fits sales almost as well as much more complex models that
between 0.00007 and 0.03 with higher values for Europe than
seek to correct its limitations (Chandrasekaran & Tellis, 2007).
for the U.S. (Sultan, Farley, & Lehmann, 1990). Values for the
Consistently with these results, the simple Bass model still domi-
parameter q would be in the interval from 0.38 to 0.53. How-
nates other approaches in many areas of application, and its
ever, these intervals are too large to provide a useful guidance to
parsimony and good face value could durably foster its success.
practitioners, except for providing references for reasonableness
Counteracting to these advantages, the main critics formu-
checking.
lated to the Bass approach are related to the absence in the
Proceeding with sources that provide point estimate, we find
model of other explanatory variables like prices or advertising
for instance Becker, Siduh, and Tenderich (2009). These authors
efforts. This gave rise to the Generalized Bass model which was
apply in a simulation study values of p ¼ 0.01, 0.02 or 0.025
shown, under fairly general assumptions, to collapse to the
corresponding to three different scenarios, and a value of q ¼ 0.4
simple Bass model. Other critics relate to the lack of micro
for all three scenarios. M is assumed to be a value of 70% or 90%
foundation of the model: Actually the model contains no positive
percent of the light-vehicle market volume of one year. However,
description of how individuals behave. The assumption about
the justifications for these quantifications seem weak at least;
innovators and imitators provides only limited insights into the
they just rely on the interval of values found, for a series of
mechanisms of the choice processes of individuals. Additionally,
technologies, referring to (Mahajan et al., 1990)2. Other authori-
it remains unclear whether innovativeness and imitativeness are
tative sources are exemplified in Gross (2008) who assumes the
a priori features of an individual (potentially defined in relation
Bass parameter values p ¼ 0.01 and q ¼ 0.1. In direct email-
to the good at hand) or are a posteriori clusterings of individuals
communications, this author states that he relied on sugges-
based on observed behavior (explained by more complex
tions by Bass (2004) und Schneider (2002) to select these values:
mechanisms). In this latest assumption, it would be the explo-
The first reference (Bass, 2004) is fairly general, while the second
ration of these more profound mechanisms (and we touch here
one (Schneider, 2002) does not appear accessible to the research
the realm of consumer psychology) that would be of ultimate
community. Other authors (Davidson, Cross-Call, Craig, & Bhar-
interest for the understanding and forecast of market diffusion.
atkumar, 2013) select p and q values “based on established values”
In other words, the Bass model may describe diffusion processes,
but eventually refer to Becker et al. (2009); the market potential
but it may not explain them. But eventually it is this explanation
M is assumed equal to the number of households in the studied
that would be necessary to grant the approach sufficient realism
area multiplied by “0.03, 0.25, and 0.7 in the low, medium, and high
and versatility.
growth scenarios”. All in all, these literature sources provide only a
limited contribution for practitioners looking for reliable Bass
2.3. The critical role of Bass parameters
diffusion parameters. Additionally, the analyzed literature does
not provide strong support for its assumption: quoted sources
In a context where the Bass model has become very influential,
appear evanescent or very general with, additionally, some
it is important that sufficient confidence can be obtained in the
tunneling effect when an author quotes an author who quotes an
realism of the various parameters that ultimately drive the
author whose paper is not available. It may then be that sec-
outcome of the method. There are, first, some issues related to the
ondary sources of data may not be helpful for the potential users
estimate of M, see (Massiani, 2012). In circumstances where,
of the Bass approach. Observing that parameter values based on
instead, the analyst can rely on sound estimates of the market
(at least apparently) authoritative sources provide limited help;
potential (this can be for instance provided by the application of a
this argues in favor of own ad hoc estimates of Bass p and q
discrete choice model as in Massiani (2015)), analysts can focus on
parameters.
the values to choose for p and q. While the literature proposes a
number of assumptions or estimates for these values, to the best of
our knowledge, no critical review seems available that investigates 2
The reference section of the cited paper (Becker et al., 2009, p.31) does not
how and whether practitioners may find a practical solution to the provide information about this entry. Even after making additional research, it does
proper quantification of p and q. For research about the diffusion not appear possible to provide accurate information on this source.
20 J. Massiani, A. Gohs / Research in Transportation Economics 50 (2015) 17e28

3.2. Ad hoc estimates in the literature considering multiple units ownership of automobiles in Australia
(1966e1996). The reported p and q values for the first car pur-
Sources that perform proper estimates can be distinguished chase models are p ¼ 0.0076 and q ¼ 0.090 (Steffens, 2003), with q
based on whether they went through a review process or not. (with his notations, proportion of the population that can adopt
While contemporary science puts a high premium on “reviewed” the innovation) equal to 0.914. The limitation of this estimate, for
papers, we believe that, especially in the Bass research area, un- other purposes, is that they relate to market conditions (time
reviewed works deserve attention.3 framework, location) that cannot easily be transferred to other
Unreviewed works include for instance the estimation by Mac markets.
Manus and Senter on HEV annual sale volumes in the U.S. be- Other authors estimate parameters for one market to apply
tween 1999 and 2008 (McManus & Senter, 2009). Based on their them for estimations for another market. This can relate to tech-
data set, and an assumption of a 1.9 million vehicles' market po- nologies (for instance coefficients estimated from Hybrid sale
tential, the authors estimate the Bass coefficients pair 0.0026 and volumes are applied to forecast the market diffusion of Battery
0.709. Additionally, the authors provide other estimates for a Electric Vehicles), but also geographically (from one country to
generalized Bass model that includes the fuel price as an addi- another). This second situation is illustrated by Park, Kim, and Lee
tional explaining variable. Additional to Mac Manus and Senter, a (2011) who estimate model parameters p, q, m for Korea based
number of estimates are available in Master Theses that we detail on HEV sales in Japan between 1997 and 2006. Subsequently,
in footnote.4 they “convert” the estimated imitation factor to the Korean context

Model Potential (Vehicles) p q R2 p-value

Toyota Prius 2 875 826 0.001645368 1.455182 0.962873 0.001

Civic Hybrid 3 685 991 0.003437559 0.631287 0.956147 0.001
Ford Escape 368 377 0.03673027 0.432231 0.999401 0.001

Interestingly, most of the unreviewed papers we found relate to by incorporating results from a comparison of imitation parame-
studies about the energy sector. We interpret this as a signal about ters across several Pacific Rim countries (including Japan and Ko-
the attractiveness of the Bass model for non-economists, which is rea) for previously launched products (based on Takada & Jain,
due to its good face value. In the next paragraphs, it is considered 1991). In their study, the innovation factor remains unchanged,
whether this situation changes if reviewed papers are now because in the author's words, “the estimated value for the imitation
considered. factor has to be adjusted because the innovation factor reflects the
Starting from products that only partly relate to new automo- characteristics of the product and/or technology itself, but the
tive solutions, Steffens (2003) suggests an extended Bass model by imitation factor reflects not only the product or technology charac-
teristics but also customer characteristics (Bass, 1969)“5 They obtain
(p; q) values equal to (0.0037; 0.3454) and a potential of 10.2
3
million vehicles. A generalized Bass model is also calibrated using
They testimony on how Bass models are used in real world applications. In the
price data relating to the relative price of “Prius” and its Internal
context of Electric Vehicles where time series of sales data are currently available
only for a few years, important information could be present in unreviewed texts. Combustion Engine (ICE) equivalent car “Corolla”, from 1997 to
4
We posit that Master theses (which could be considered as reviewed or not 2006. Other authors use numbers of new EV car registrations in
reviewed) can provide insights in the implementation of Bass models in applied Norway (Jan. 2003 until June 2013) and implement them in a
studies. Master theses reflect the ability of an educational system to impart model for Denmark (Jensen, Cherchi, Mabit, & Ortúzar, 2014). This
knowledge to prospective users of the respective method. We found the following
sources:Shoemaker (2012) analyses the sales of 585 models of passenger vehicles
choice relates to the fact that EVs are still too scarce in Denmark
and utility vehicles between Jan. 1997 and Aug. 2011 on the U.S. market. He finds, while they are more abundant in Norway (see, in this issue, Fig-
that Alternative Fuel Vehicles (AFVs) have already peaked during the estimation enbaum et al.). They obtain the pair of values (0.002; 0.23) while
period. This corresponds to the observed diffusion pattern at that time where sales the assumed market potential results from the application of a
exhibited a decrease after peaking in 2007. Under these circumstances, it could be
Discrete Choice Model (a procedure also implemented in Massiani
claimed that the model estimate performs well, since the available data at that time
described a large part of the product history. But in practitioners view, the (2015)). However, a critical point could be that, if EVs are more
contention that AFV have already experienced their peak could be not convincing diffused in Norway than in Denmark, then (unless one can
(although conform to the episode observed in the time frame considered by establish that the differences in market development are only due
Shoemaker). Additionally, Shoemaker performs an out of sample forecast test, that to difference in supply) there can be something different in the
provides moderate adherence to the out of sample observations. This is one of the
demand in one country compared with the other. This difference
few cases where the forecast performance of Bass models is evaluated in the
literature and these results provide a reason to consider Bass driven results with could make it illegitimate to transfer parameters. Concerning the
additional attention.Other non referred ad hoc estimates relate to an application to geographical issue, as well as to the technology transfer issue, we
a “Beijing EV recharging simulation” proposed by Li, G. (nd). Power Forecasting for suggest to undertake further research activities about the perfor-
Plug-in Electric Vehicles with Statistic Simulations. The author estimates Bass pa-
mance of Bass-models relying on transferred parameter values.
rameters for sales volumes. However, the applied method is not thoroughly
described and remains unclear. Li obscurely refers to a parsimonious description of Finally, in some studies, parameters are calibrated on data from
a “statistic model of EV's in US from 1999 to 2008”. Assuming an exogenous po- the same markets than the ones of application. So the researcher
tential of 1.9 million sales of EV's, the author obtains the values p ¼ 0.0000365 and was in the fortunate position to have sales data at hand for the
q ¼ 0.447.A study by Cordill is based on sale volumes for various Hybrid models market for which his research was conducted. One of these
between 2000 and 2010 in the U.S., and a forecast of the Plug-in Hybrid Electric
Vehicle diffusion is provided. Cordill, A. (2012). Development of a diffusion model
to study the greater PEV market. A Thesis Presented to The Graduate Faculty of The
University of Akron In Partial Fulfillment of the Requirements for the Degree Master 5
Consultation of the original Bass paper does not provide, to our view, strong
of Science.Cordill (2012) reports the following Bass parameter values. support to this statement.
 J. Massiani, A. Gohs / Research in Transportation Economics 50 (2015) 17e28 21

Table 1
Coefficient values for a Bass and a Gompertz diffusion model (monthly data and annual approximation, Lamberson, 2008).

Bass Innovation coefficient p Imitation coefficient q Predicted total market potential M

0.0000515 (monthly) 0.0728 (monthly) 1 600 000 (monthly)
0.000618 (approx. annual) 0.8736 (approx. a.) 19 200 000 (approx. a.)
Gompertz Coefficient a Coefficient b Predicted total market M
11.2 (monthly) 0.0118 (monthly) 25 700 000 (monthly)

Table 2
Bass coefficient values estimated by Cao (2004) for innovative automotive technologies.

Innovation coefficient p Imitation coefficient q Predicted total market potential M

Bioethanol (p.66) 0.00441 0.491 245 971

CNG (p.58) 0.021 0.265 100 371
Hybrid (p.70) 0.000446 0.4788 Exogenous assumption: 10% of total registered cars in 2000

contributions was made by Lamberson (2008) on Hybrid Electric and a strong imitation effect. This will typically result in a diffusion
Vehicles. Lamberson's final aim was to forecast the diffusion of curve that will have a lengthy start-up phase and a steep increase of
Hybrid Electric Vehicles in the USA based on monthly vehicle diffusion in a second phase.
registrations from Feb. 2001 to Oct. 2007 (Lamberson, 2008). Pa- It could be hypothesised that the very low p value for Hybrid
rameters of a Bass model and a Gompertz model were estimated by results from the shortness of the time series available at that time.
the method of non-linear least squares. Table 1 provides the results The author (Cao, 2004, p. 68) reports: “Honda Insight, the first HEV
together with a conversion to approximated annualized parame- model in the U.S., was introduced in December 1999. Thus the number
ters values for comparison purposes. of observations on HEV sales is even smaller than that for CNG and E85
Lamberson supports a low innovation coefficient. This result is [consisting of 85% ethanol and 15 % fossil fuel] vehicle sales. Currently,
consistent with a detailed study made by Cao (2004). Cao provides our HEV data contain the annual sales for only four consecutive
estimates for sales of Alternative Fuel Vehicles in California. He years: 2000e2003.” In particular, one could doubt that the diffu-
made two general assumptions to provide guidance to the inter- sion of the Hybrid technology had reached its inflexion point at that
pretation of his results: time, casting doubt on the whole estimation. On this issue, we
report the assessment made by Cao:
(1) “We relaxed one major assumption underlying the basic
“( … ) the estimates of these parameters [p;q;M] are sensitive to
first-purchase Bass model d we assumed that the market
the number of observations available to the estimation
potentials of AFVs [Alternative Fuel Vehicles] do not keep
(Mahajan et al., 1990). Heeler and Hustad (1980) suggested that
constant but vary as explanatory variables change.” (Cao,
stable and robust parameter estimates can be obtained only if
2004, p. 58).
the data under consideration contain at least ten observations
(2) “We assumed that the market potential of HEVs [Hybrid
and include the peak of the non-cumulative adoption curve,
Electric Vehicles] is around 10% of total car and truck regis-
which is supported by the results of Srinivasan and Mason
trations in 2000.” (Cao, 2004, p. 68). Cao derives the 10%
(1986).”
assumption from an EIA scenario, in which 19 million Hybrid
Electric Vehicles sales between 2001 and 2025 are compared
to 220 million vehicles registered in 2002. So 10% roughly Cao pursues that waiting to have enough observations may just
equates to the share of the current registered fleet compared provide results “too late to use the estimates for forecasting purposes”
to total sales in the period 2001e2015 (see p.68).6 In sub- (Mahajan et al., 1990 p. 9), making the prediction useless (Hyman,
sequent years, the potential is supposed to vary as a function 1988). “Caught by this fundamental paradox, most operationally
of awareness about the existence of HEV and (lagged) fuel useful (as opposed to post-hoc diagnostic) diffusion models are cali-
price. In other words, in this model, the market potential is brated on relatively little data and hence are not necessarily stable.
supposed to vary at each time step, starting from 10% of total However, this is a virtually inescapable feature of this modeling
registrations, as a function of “awareness” and lagged fuel approach.” (Cao, 2004).
price.

Cao's estimates are shown in Table 2 below, where results for 3.3. Summary of available estimates
“Hybrid” refer to model 2 which Cao has chosen as his “final best
model”. In Table 3, we provide a summary of the main results about Bass
These values are fairly untypical compared with values cited in coefficients for new propulsion technologies which we found in the
the literature for other products. As noted by Cao, the estimated literature. As these elements suggest, there is a wide discrepancy
innovation coefficient varies strongly when the initial market among p and q values for modeling the market diffusion. A pro-
diffusion assumption changes. We observe that the coefficient spective user of the Bass diffusion model would probably be in
values estimated by Cao imply a very weak innovation for hybrids difficulty when looking for evidence based values of the
coefficients.
The question that arises then is whether we get better insights
about plausible p and q values by estimating p and q from observed
6
Cao, 2004 does not precisely provide reasons for this choice. market sales. In the next section, we present our own estimations
22 J. Massiani, A. Gohs / Research in Transportation Economics 50 (2015) 17e28

Table 3
Available literature estimates for p and q parameters.

Source technology Method Innovation Imitation Market potential

coefficient p coefficient q

Becker, Sidhu Authoritative sources 0.01 0.4 Exogenous: 70% or 90% of the light-vehicle
et al. (2009) 0.02 market in each year.
Electric cars 0.025
Davidson et al. (2013) Authoritative sources, but eventually refers to Idem Idem Exogenous: Number of Household are “0.03,
Electric cars Becker et al. (2009) 0.25, and 0.7 in the low, medium, and high
growth scenarios”
Gross (2008) Authoritative sources 0.01 0.1
Li (nd) Based on “a statistic model of EVs in US from 0.0000365 0.447 Exogenous: “At most half of the vehicles in
Electric cars 1999 to 2008” market can be EVs, the ultimate market
potential is calculated as 2.5 million”.
Cordill (2012) Calibration on US market data 2000e2010 Estimated
Prius Hybrid 0.0016 1.45 2.87 million
Civic Hybrid 0.00343 0.631 3.68 million
Ford Escape Hybrid 0.036 0.432 0.36 million
Steffens (2003) First car purchase (not specifically AFV) in Australia 0.0076 0.0905 Exogenous: 91% of the population would finally
Conventional cars 1966e1996 purchase the technology.
Shoemaker (2012) a Calibration on AFV monthly sales in the US Dec Estimated
Passenger vehicles 1995 (Oct 2004 for utility AFV) Dec 2011 in the US 0.0912 0.4692 436 000
Utility vehicles 0.008124 0.4632 2 300 000
Lamberson (2008) a Calibration on US monthly sales Estimated
HEV (Feb. 2001eOct. 2007) 0.000618 0.8736 1.6 million veh.
Park et al. (2011) HEV sales in Japan (1997e2006) parameterized to
HEV Koreab 0.0037 0.3454 10,2 million veh.
Jensen, CherchiMabit, Norwegian new electric car registration data from Exogenous: Result of a Discrete Choice Model
and Ortúzar (2014) (Jan. 2003eJun. 2013)
Electric cars 0.002 0.23
McManus and HEV annual Data in the US. (1999e2008) 0.0026 0.709 Estimated: 1.9 million veh.
Senter (2009)
HEV (0.00124 (0.77922
generalized Bass) generalized bass)
Cao (2004) Calibration on annual sales in the US 1993e2002
E85 0.00441 0.491 E85: 245971
CNG 0.0210 0.265 CNG: 100000
Hybrids 0.000446 0.4788 Hybrid: Exogenous. Based on EIA scenario of
19 million HEV sold until 2025, subsequently
varies in function of HEV awareness and
(lagged) fuel price
a
Monthly coefficients are annualized by multiplication with factor 12.
b
Imitation factor modified based on comparison of parameters across several Pacific Rim countries for previously introduced products (see Takada, H. and D. Jain (1991).
“Cross-National Analysis of Diffusion of Consumer Durable Goods in Pacific Rim Countries.” Journal of Marketing 55(2): 48e54.)

based on data about the real diffusion of various alternative pro-
Yearly data: New registrations for LPG, CNG, EV and HEV tech-
pulsion technologies in Germany. nologies are available for the period 2005e2013 (EV since
2004).
4. Estimations based on the German market data
These data constitute the most recent available at the time the
estimation is performed. The relatively large quantity of available
In this section, we report Bass parameter estimates, based on
data may help providing reliable estimates.
real market data in Germany. In the next subsection, our data set is
To estimate Bass parameters, the cumulative new vehicles'
described. Then, the estimation process and results are presented
registrations Nt of a given technology at time t have to be used. For
considering two alternative computation methods: In the first, the
this purpose, an estimate of the stock of vehicles on Dec. 31, 2008 is
market potential is endogenous in the estimation. In the second,
available, based on annual new registrations since 2005 (Electric:
the market potential is treated as an exogenous variable in the
since 2004).7 The evolution of vehicle registrations and cumulative
model.
registration is provided in Fig. 2 and Fig. 3 respectively. For the
estimation of the Bass parameters, these time series of cumulative
4.1. Available data new registrations Nt are matched with the new registrations nt in
the same time frame.
The estimation is based on time series of sales volumes of a set Based on these data, Bass parameter values are estimated using
of innovative technologies. As far as the German car market is two different fashions, depending on whether the parameter M
concerned, the Kraftfahrtbundesamt (KBA) provides monthly and (the market potential) is endogenous in the estimation (this is the
yearly data on registrations for new cars differentiated by situation commented by Cao (2004)) or whether it is determined
technology. exogenously. So, in synthesis, the alternatives are as follows:

Monthly data: New registrations for Liquefied Petroleum Gas

(LPG), Compressed Natural Gas (CNG), Electric Vehicles (EV) and
Hybrid Electric Vehicles (HEV) are available from Jan. 2009 to 7
It has been checked that monthly and annual data are compatible for the years
Sept. 2014. 2009e2013, so the correspondence of both time series is perfect.
 J. Massiani, A. Gohs / Research in Transportation Economics 50 (2015) 17e28 23

3000

2500

2000

1500

1000

500

2012.12

2013.06
2008.12
2009.03
2009.06
2009.09
2009.12
2010.03
2010.06
2010.09
2010.12
2011.03
2011.06
2011.09
2011.12
2012.03
2012.06
2012.09

2013.03

2013.09
2013.12
2014.03
2014.06
2014.09
LPG (incl. bivalent) new registr. CNG (incl. bivalent) new registr.
Electric new registr. Hybrid new registr.

Fig. 2. Monthly passenger car new registrations in Germany of various alternative propulsion technologies.

140000

120000

100000

80000

60000

40000

20000

0
2009.06

2013.09
2012.12
2008.12
2009.03

2009.09

2011.12
2009.12

2013.12
2010.03
2010.06
2010.09
2010.12
2011.03
2011.06
2011.09

2012.03
2012.06
2012.09

2013.03
2013.06

2014.03
2014.06
2014.09

LPG (incl. bivalent) cumulaƟve registr. CNG (incl. bivalent) cumulaƟve registr.
Electric cumulaƟve registr. Hybrid cumulaƟve registr.

Fig. 3. Cumulated monthly passenger car new registrations in Germany of various alternative propulsion technologies.

1. Estimation of p, q and M based on the sales of the technology in Starting with the equation of the Bass diffusion model, it is possible
the first years of diffusion. In this estimation process, the market to make some transformations useful for estimation such that:
potential is endogenously estimated together with Bass pa- dNt q
rameters p and q. nt ¼ ¼ pM  pNt þ qNt  Nt2 (4)
dt M
2. Estimation of p and q based on the sales and an exogenous
potential M, as assumed by the modeler based on other And, in a second step:
information.

dNt q
nt ¼ ¼ pM þ ðq  pÞNt  Nt2 (5)
dt M

4.2. Estimation with model-endogenous market potential

Or
First, the coefficients p and q are estimated together with the dNt
market potential M of a technology. The dependent variable is new nt ¼ ¼ b0  b1 Nt þ b2 Nt2 (6)
dt
registrations nt, the explanatory variable is cumulative purchases
q
Nt until period t. with b0 ¼ pM, b1 ¼ qp, and b2 ¼ M .
24 J. Massiani, A. Gohs / Research in Transportation Economics 50 (2015) 17e28

Table 4
Estimated Bass-parameters for annual registrations.

Technology Estimated regression-coefficients, (adj.) R2 and radicand Bass-parameters

b0 b1 b2 R2 adj. R2 Radicand p q M

Electric 53 1.249 4.41E-05 0.96 0.94 1.570 0.0019 1.2513 28 383

Hybrid 4795 0.131 2.17E-06 0.96 0.94 2.44E-02 NA NA NA
LPG 5885 0.294 4.92E-06 0.21 0.10 0.202199 0.0779 0.3718 75 566
CNG 12 972 0.084 3.19E-07 0.68 0.55 2.36E-02 0.1187 0.0349 109 276

Based on: KBA, Annual registrations (2006e2013, Electric: 2005e2013) and cumulated new registrations since 2005 (Electric: since 2004) in Germany.

Table 5
Estimated Bass-parameters for monthly registrations.

Technology Estimated regression-coefficients, (adj.) R2 and radicand Bass-parameters

b0 b1 b2 R2 adj. R2 Radicand p q M

Electric 11 0.071 1.74E-06 0.84 0.84 5.16E-03 0.00027 0.07159 41 236

Hybrid 463 0.045 1.72E-07 0.80 0.80 1.68E-03 0.00186 0.04281 249 452
LPG 1400 0.024 1.68E-07 0.14 0.11 3.83E-04 NA NA NA
CNG 5545 0.159 1.24E-06 0.45 0.43 2.23E-03 NA NA NA

Based on: KBA, Monthly registrations (Jan. 2009eSept. 2014) and cumulated new registrations since 2005 (Electric: since 2004) in Germany.

If a (real) solution to the polynomial equation exists in real 1986). Using different initial values is also highly recommended
numbers, once b0 to b2 are estimated, the values p, q and m can be in practice (Putsis & Srinivasan, 2000).
calculated.8 In our study, the Bass parameters p, q and M are Comparing both techniques to estimate Bass coefficients from
calculated from the OLS estimated regression coefficients b0, b1 and the differential equation (3), the NLS approach is seemingly
b2 of equation (6), which represents a transformation of the dif- preferred to the MLE in the literature: (Srinivasan & Mason, 1986, p.
ferential equation (3). 178) claim that, Schmittlein and Mahajan consider in their MLE
Other authors recommend to estimate the Bass coefficients formulation “only sampling errors and ignore all other errors, such as
directly from the differential equation (3). Since the differential the effects of excluded variables and the misspecification of the
equation (3) is non-linear, the OLS technique is not applicable. probability density function for adoption time”, and thus are likely to
Alternatively, MLE and NLS techniques are proposed to estimate the underestimate the standard errors of these parameters. Mahajan
parameters directly from the differential equation of the diffusion and Sharma compared both estimation techniques empirically and
model (Schmittlein & Mahajan, 1982; Srinivasan & Mason, 1986). found an overall superiority of the NLS procedure (Mahajan &
This topic is discussed extensively in the literature for instance in Sharma, 1986). A simulation study by Srinivasan and Mason
Jiang et al. (2006); Boswijk and Franses (2005); Van Den Bulte and (1986) demonstrated that the biases in the parameter estimates
Stremersch (2004); Venkatesan, Krishnan and Kumar (2004); are small, less than 7%. More recently, Van den Bulte and Lilien
Lilien, Rangaswamy, and Van den Bulte (2000); Sultan et al. (1997a, 1997b) stated that the estimators in the NLS estimation
(1990); and Van den Bulte and Lilien (1997a, 1997b). procedure are consistent but not unbiased. Specifically, by exam-
According to Srinivasan and Mason, the model formulation of ining both empirical and simulation data, they found that NLS tends
the NLS approach is as follows: to underestimate m and p and overestimate q to a much greater
extent than stated by Srinivasan and Mason (1986). They pointed
out that the magnitude of the bias depends on the amount of noise
SðtÞ ¼ m½FðtÞ  Fðt  1Þ þ ε ðtÞ (7) in the data, the number of observations, and the differences be-
Both techniques (NLS and MLE) provide standard errors of the tween the cumulative penetration of the last observation and the
parameters. However, they have their own limitations. For both true saturation penetration. Thus, one possible solution to mini-
approaches, an initial value for each parameter is required to esti- mize the bias is to increase the number of data points, which is
mate the Bass model and the parameter estimates can be sensitive consistent with the findings of Putsis (1996).
to the initial values. The model estimation may not converge in Generally, the superiority of estimating the Bass coefficients
some cases, partly due to poor initial values (Judge, Griffiths, Hill, directly from the differential equation (3) rather than from its
Lütkepohl, & Lee, 1986). Thus OLS estimates obtained from esti- transformation (6) is not established. For this reason, the presented
mation of the transformed equation (6) could be used to provide results are based on the OLS estimate of its transformation (equa-
the initial values for these parameters (cf. Srinivasan & Mason, tion (6)). The results are presented in Tables 4 and 5.
Available data indicate that there were about 71 545 LPG new
registrations in 2005e2014 and a fleet stock for LPG of 500 867 (Jan.
8 b0 b2
b0b2 ¼ qp and division by q yields: q ¼ p (1). Insert (1) in eb1 ¼ qp,
yields eb1 ¼ q þ b0qb2 , and addition of b1 yields: q þ b1 þ b0qb2 ¼ 0. Multiplication by
q yields: q2þb1qþb0b2 ¼ 0. This second order polynomial has the following roots: Table 6
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 Estimated Bass-parameters for electric vehicles (endogenous potential estimate).
q1=2 ¼ b21 ± b41  ðb0 b2 Þ The index on q indicates that there are possibly up to
two solutions in real numbers. Mostly we receive one positive and one negative q Frequency of time series p q M
value, but only positive values can be meaningful interpreted in the framework of
Annual data 0.0019 1.2513 28 383
the Bass model. In this case the only positive value is reported. If both solutions are
Monthly dataa 0,00322 0,85912 41 236
negative, the higher value is reported. A mathematical solution in real numbers
a
exists if the radicand is not negative. If the radicand is negative, p, q and m values Annualized p and q parameter values (approximated by multiplication of
cannot be calculated. In this case the term “NA” (Not available) is reported. parameter values for monthly data by 12).
 J. Massiani, A. Gohs / Research in Transportation Economics 50 (2015) 17e28 25

Table 7
Estimated Bass p values (monthly new registrations) for different assumed market potentials.

Bass-p-values Market potential in million passenger cars

Technology 0.1 0.5 0.7 1 5 10 20

Electric 3.56E-04 9.26E-05 6.71E-05 4.75E-05 9.69E-06 4.86E-06 2.43E-06
Hybrid 2.86E-03 1.10E-04 3.40E-05 7.50E-05 3.12E-05 1.65E-05 8.46E-06
LPG 7.99E-03 2.10E-03 1.49E-03 1.04E-03 2.05E-04 1.02E-04 5.10E-05
CNG 7.11E-04 1.49E-03 1.05E-03 7.31E-04 1.44E-04 7.16E-05 3.58E-05

Data KBA: Registrations (Jan. 2009eSept. 2014) and cumulated new registrations since 2005 (Electric: since 2004) in Germany.

1st, 2014). So the estimated market potentials reported in Table 4 Table 8

are unrealisticly low for the LPG technology. For Electric Vehicles Estimated Bass q values (monthly new registrations) for different assumed market
potentials.
a q-coefficient larger than one is obtained. For Hybrids, there exists
no (real) solution for the p and q, because the square root of the Bass-q-values Market potential in million passenger cars
radicand is negative. The results reported in Table 4 are estimated Technology 0.1 0.5 0.7 1 5 10 20
from annual data. Table 5 provides estimates from monthly data. Electric 0.055 0.048 0.047 0.047 0.046 0.046 0.046
From estimations of monthly data we get negative radicands for the Hybrid 0.036 0.029 0.027 0.025 0.022 0.022 0.022
LPG 0.012 0.006 0.006 0.007 0.007 0.007 0.007
LPG and CNG technologies, so that no valid p and q values can be
CNG 0.025 0.001 0.001 0.002 0.002 0.002 0.002
calculated. For Hybrids, a negative p value is calculated. Only for the
Data KBA: Registrations (Jan. 2009eSept. 2014) and cumulated new registrations
technology “Electric” we derive plausible p- and q-values but a low
since 2005 (Electric: since 2004) in Germany.
estimate for the market potential. Since p- and q-values are esti-
mated for monthly periods we multiply these values to approxi-
mate annual p and q values, see Table 6.
4.3.1. Model based on monthly data for Electric Vehicles, Hybrids,
So the ad hoc estimated parameters are of limited benefit for
LPG and CNG
practitioners, at least considering the data currently available. In
First, estimates are made based on monthly data from KBA on
order to further check for the possibility that a particular data
new registrations (Jan. 2009eSept. 2014) and cumulative new
configuration may lead to this result, estimations have been applied
registrations since 2005 for EV, HEV, LPG and CNG. Additionally, in
to another data set relating to sales in Japan and overseas. These
order to investigate the sensitivity of the p and q estimates to the
other estimates, presented in the appendix lead to a similar
market potential, various values for market potentials are consid-
conclusion. This pessimistic conclusion echoes certain findings in
ered: M ¼ 0.1, 0.5, 0.7, 1, 5, 10 and 20 million vehicles. The resulting
the literature: Van den Bulte and Lilien (1997a, 1997b p. 2045) claim
values of p and q parameters are provided in Tables 7 and 8.
that in application of the model, “the estimated population size
The results of the estimations are, of course, dependent on the
[meaning market potential M] is close to the cumulative number of
assumed market potential for various technologies. The market
adopters observed in the last time period for which data are
potential of 10 million cars is arguably a reasonable estimate for the
available.” They also claim that this approach results in an “upward
size of the market potential M for Hybrid Electric Vehicles9. The
bias in the contagion [imitation] parameter.”
estimated parameters in Tables 7 and 8 correspond to monthly
This, pessimistic, conclusion could however result from the
values, the corresponding annual p and q values are approximated
endogenous approach at stake. The next section proceeds with the
by multiplication with the factor 12. Accordingly, an annual p-
other approach, the one relying on the exogenous quantification of
value of 0.0002 and an annual q-value of 0.26 constitute the best
the market potential M, thus focusing on the sole estimate of pa-
available estimates for Hybrids' diffusion.
rameters p and q. In such a situation, M is defined a priori, inde-
pendently of the estimate of p and q.
4.3.2. Model based on annual data for Electric Vehicles, Hybrids,
LPG and CNG
4.3. Estimation with exogenous market potential Recalling that KBA data for new propulsion technologies are
also available on an annual basis for the period 2005e2013, p and q
In this section, the market potential M of the Bass-model is fixed are also fitted on annual sales data. Additionally, to complement
exogenously. When calibrating p and q based on exogenous po- these models, GDP growth rates and a dummy variable for car
tential assumptions, much attention should be dedicated to the scrappage scheme (“Abwrackpra €mie”) were tested as explanatory
quantification of the potential M; especially the estimate of the Bass variables in the models. Since these variables were not found sig-
p parameter depends on the a priori assumed market potential, see nificant or did not have additional explanatory power, they were
Table 7. The quantification of M is not an easy task. This is illustrated discarded.
by the general lack of solid argumentation on the value chosen for These results (Tables 9 and 10) provide estimates of the correct
the parameter M that can be observed in the existing literature. sign for Electric Vehicles and LPGs, and for realistic market poten-
The estimation of Bass parameter values proceeds as follows, tial numbers also for Hybrids. For CNG we do not get realistic, i.e.
starting again from equation (1).
This equation can be rewritten as:
9
This estimate is derived from a metaanalysis based on a Synthetic Utility
dNt Function that produces forecasts of market shares for competing propulsion tech-
nt ¼ ¼ pXt þ qYt (8) nologies, see Massiani, J. (2012). Using Stated Preferences to forecast alternative fuel
dt
vehicles market diffusion. Italian Journal of Regional Sciences. 11: 93e122. This
1 N ðM  N Þ. latest model suggests an average market share of hybrid cars over the next ten
with Xt ¼ MNt and Yt ¼ M t t years of 25% (before diffusion is accounted for). Applying this 25% estimate to the
If M and Nt are given, both Xt and Yt can be calculated. This stock of cars in Germany of around 40 million suggests a market potential of 10
makes it possible to estimate p and q. million cars.
26 J. Massiani, A. Gohs / Research in Transportation Economics 50 (2015) 17e28

Table 9
Estimated Bass p values (annual new registrations) for different assumed market potentials.

Market potential in million passenger cars

0.1 0.5 0.7 1 5 10 20

Electric 1.11E-03 2.53E-04 1.82E-04 1.28E-04 2.60E-05 1.30E-05 6.50E-06
Hybrid 2.15E-02 4.37E-03 3.38E-03 2.49E-03 5.42E-04 2.74E-04 1.37E-04
LPG 6.94E-02 1.61E-02 1.15E-02 8.08E-03 1.62E-03 8.12E-04 4.06E-04
CNG 1.26E-01 2.70E-02 1.92E-02 1.34E-02 2.67E-03 1.33E-03 6.66E-04

Data KBA: registrations (2006e2013, Electric: 2005e2013) and cumulated new registrations since 2005 (Electric: since 2004).

Table 10
Estimated Bass q values (annual new registrations) for different assumed market
potentials.

Market potential in million passenger cars

0.1 0.5 0.7 1 5 10 20

Electric 1.061 1.010 1.007 1.004 1.000 0.999 0.999
Hybrid 0.789 0.357 0.339 0.327 0.305 0.302 0.301
LPG 0.204 0.024 0.018 0.013 0.005 0.004 0.003
CNG 0.071 0.097 0.101 0.103 0.108 0.108 0.109

Data KBA: Annual registrations in Germany (2006 e 2013, Electric: 2005e2013) and
cumulated new registrations since 2005 (Electric: since 2004).

positive q coefficients for market potentials between 0.5 and 20

million cars. The q values close to one for EV are not reasonable. For Fig. 4. Scatter plot of available p and q values.
LPG, Bass q parameter values decline with a growing market po-
tential and are unrealistically low. Control variables for GDP in
levels and returns, nominal and real were tested but are not sig- results in a bisection of the Bass p value. This high sensitivity makes
nificant. Altogether, given an exogenous market potential, only for it very difficult to draw conclusions for the actual value of p.
Hybrids we get plausible Bass estimates from monthly as well as Interestingly, the estimated Bass q parameter values vary only
annual time series. For Electric Vehicles we get from monthly time marginally for this range of market potentials. Since we believe that
series results which are consistent with the Bass approach. The for car technologies the actual market potential should fall in this
helpfulness of these estimates need however to be assessed taking range of assumed market potentials or could just be higher, we
in consideration the other results obtained. recommend to derive the Bass q parameter value from an estima-
Additionally, it becomes apparent that the Bass p value is sen- tion with exogenous market potential. If a Bass q value is estimated
sitive to the assumed market potential. For a wide range of which is not realistic, this could be that the diffusion process may
hypothesised market potentials m between 0.5 and 20 million not follow Bass or that the time series available for estimation
passenger cars, we find an inverse relationship between the purposes is too short and does not contain enough information. So
assumed m and the estimated p value. Precisely, a doubling of m

Table 11
Comparison of Bass parameters estimates.

Source Innovation coefficient p Imitation coefficient q

1. Literature estimates
Becker et al. (2009) 0.01, 0.02 or 0.025 0.4
Davidson et al. (2013) Idem Idem
Gross (2008) 0.01 0.1
Li (nd) 0.0000365 0.447
Cordill (2012) Prius 0.0016 1.45
Hybrid Civic 0.00343 0.631
Ford Escape 0.036 0.432
Steffens (2003): Conventional cars 0.0076 0.0905
Schoemaker (2012)* 0.0912 (pass. veh) 0.008124 (utility veh) 0.4692 (pass.veh) 0.4632 (utility veh.)
Lamberson (2008) HEV* 0.000618 0.8736
Park et al. (2011) 0.0037 0.3454
Jensen et al. (2014) 0.002 0.23
McManus and Senter (2009) 0.0026 0.709
Cao (2004) e Hybrids 0.000446 0.4788
2. Own estimates
endogenous potential e KBA yearly data 0.0019 1.2513
endogenous potential e KBA monthly data* 0.00322 0.8591
exogenous potential (¼ 10 mln) - KBA annual data 0.000013 0.999
0.000274 (Hybrids) 0.302 (Hybrids)
exogenous potential (¼ 10 mln) e KBA monthly data* 0.00005832 0.552
0.000198 (Hybrids) 0.264 (Hybrids)
Toyota Hybrid, worldwide 0.0007 0.2422
Toyota Hybrid, Europe 0.0008 0.2318

Note: * Parameters estimated based on monthly data were annualized by multiplication with factor 12.
 J. Massiani, A. Gohs / Research in Transportation Economics 50 (2015) 17e28 27

the estimate (with exogenous market potential) of the Bass q value when the potential is endogenous in the estimation), will be
could serve as a first step to test if Bass is appropriate to fit the assumed prior to application of the Bass model.
diffusion process. If the researcher wants to apply Bass despite an Third, one can observe that the two available sources (Cao and
unreliable q value from ad-hoc estimation, she could transfer Bass Lamberson) that, considering the investigated product, most
values estimated for other markets (published in the literature or closely match the German electric mobility market, exhibit
from own ad-hoc estimations). If the ad-hoc estimated Bass q value anomalous low values for the p coefficient. Prospective users of the
is realistic, the researcher could still be doubtful concerning the Bass approach should pay special attention when deciding to
sensible p value. Since our results reveal that there exists no transfer such values in their own models.
panacea to solve this problem, we suggest to follow different More generally, the strong pattern of dispersion of the reported
strategies: As far as the researcher is confident concerning in her parameter values suggests that the use of any of the estimated
selected exogenous market potential M, the Bass p parameter can parameters (or any aggregation or averaging thereof) would require
be determined by M and q. Otherwise, if the researcher can rather attention. The various estimates provide results that sometimes
be confident in a range of plausible market values, perhaps the differ by several orders of magnitude, and the theory does not
researcher could find a solution in a simultaneous estimation of M indicate an ultimate selection criterion among such different re-
and p given q. Following this strategy, the Bass approach with sults. Until these interrogations have not received a convincing
model-endogenous M, p and q becomes obsolete. answer, there generally appears a warning to the prospective users
of Bass parameters for the electric vehicles' diffusion.
5. Conclusions
Acknowledgment
During a research project about the evaluation of policies to
promote the market diffusion of Electric Vehicles in Germany, the We thank our former colleagues from ESMT Berlin: Jens Wein-
potential for the use of the Bass diffusion model has been investi- €rg Radeke for valuable discussions.
mann, Giselmar Hemmert and Jo
gated. During this research, additional to an interrogation on the
estimation of the market potential, an investigation on the value to Appendix. Estimates based on Toyota hybrid annual sales in
be used for Bass p and q parameters has been performed. This Japan and overseas
investigation has led us to consider values obtained through ad hoc
estimates and values proposed in the literature. The main quanti- An estimation of Bass coefficients is performed using data for
tative evidences collected during this investigation are illustrated sales of Toyota Hybrid vehicles (incl. plug-in hybrid vehicles) in
in Table 11 and Fig. 4. Japan, North-America, Europe and global for the years 1997 until
This comparison makes a few findings apparent: 2013.
First, there appears a very large discrepancy among available
estimates. There is a large variety among literature estimates and 7000
there is also a large variety (and sometimes inconsistency when we 6000
found results with the wrong sign) among ad hoc estimates. Con-
cerning the ad-hoc estimates it appears that the Bass p parameter is 5000
not stable and is sensitive to the assumed market potential M. If Global
4000
there is an uncertainty concerning the hypothesised M, our esti-
Japan
mation results show that different p values could be plausible 3000
North America
which differ in factors of up to ten or even 100. In this difficult
2000 Europe
environment we propose to follow different ways, for example to
try a two-step approach: In a first step the Bass q value could be 1000
estimated from the Bass-model with exogenous determined mar-
ket potential. In a second step, the researcher has several options to 0
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013

estimate the Bass p value given M and q or e if she is insecure about

the exogenously determined M e she can simultaneously estimate Fig. 5. Cumulative Toyota sales volumes (in 1000 Hybrids) 1997e2013.
M and p given q. Additionally, the researcher can take Bass q values Source:http://corporatenews.pressroom.toyota.com/releases/
or all Bass parameters proposed in the literature. worldwideþtoyotaþhybridþsalesþtopþ6þmillion.htm, call date: Nov. 3, 2014
Second, the hypothesised market potential (when exogenous) is
found to be highly influential for the p parameter values, while the From these time series Bass p, q and M parameter values are
q value is only marginally affected by this potential. Our results estimated:

Table 12
Estimated regression coefficients and Bass' p, q and M parameters for Toyota Hybrids.

Estimated regression-coefficients. (adj.) R2 and radicand Bass-Parameters

2 2
Sales in … b0 b1 b2 R adj. R Radicand p q M
Global 31 840 0.241 5.22E-09 0.97 0.96 0.0590 0.0007 0.2422 46 365 287
Japan 7719 0.329 2.51E-08 0.96 0.96 0.1072 0.0006 0.3280 13 049 654
North America 35 228 0.234 4.87E-08 0.79 0.75 0.0618 0.0071 0.2416 4 962 585
Europe 6887 0.231 2.64E-08 0.96 0.96 0.0541 0.0008 0.2318 8 772 998

show that a decisive element for the simulation of the market Our results provide an unconvincing negative p parameter for
diffusion is constituted by the quantification of the market poten- Japan. The results for worldwide market and European market are
tial, which in most practical cases (considering the problems arising very similar, when considering the p and q parameters, and appear
28 J. Massiani, A. Gohs / Research in Transportation Economics 50 (2015) 17e28

in a reasonable range. It is however possible to object that a model Lilien, G. L., Rangaswamy, A., & Van den Bulte, C. (2000). Diffusion models: mana-
gerial applications and software. In V. Mahajan, E. Muller, & Y. Wind (Eds.), New
calibrated on worldwide or continental data is of limited relevance
product diffusion models. New York: Kluwer Academic Publishers.
for application in a single country, a situation that may often pre- Mahajan, V., Muller, E., & Bass, F. M. (1990). New product diffusion models in
sent itself to the analyst. marketing: a review and directions for research. The Journal of Marketing, 54(1),
1e26.
Mahajan, V., & Sharma, S. (1986). Simple algebraic estimation procedure for inno-
vation diffusion models of new product acceptance. Technological Forecasting
and Social Change, 30(December), 331e346.
References Massiani, J. (2012). Using stated preferences to forecast alternative fuel vehicles
market diffusion. Italian Journal of Regional Sciences, 11, 93e122.
Bass, F. M. (1969). A new product growth for model consumer durables. Manage- Massiani, J. (2015). Cost-Benefit analysis of policies for the development of electric
ment Science, 15(5), 215e227. vehicles in Germany: methods and results. Transport policy, 38, 19e26.
Bass, F. M. (2004). Comments on ‛A new product growth for model consumer du- McManus, W., & Senter, R. J. (2009). Market models for predicting PHEV adoption and
rables’: the bass model. Management Science, 50(No. 12), 1833e1840. diffusion. Technical challenges of plug-in hybrid electric vehicles and impacts to
Bass, F. M., Krishnan, T. V., & Jain, D. C. (1994). Why the bass model fits without the U.S. power system, Subcontract No. 46827, Task 2c: Final report.
decision variables. Marketing Science, 13(3), 203e223. Park, S. Y., Kim, J. W., & Lee, D. H. (2011). Development of a market penetration
Becker, T. A., Siduh, I., & Tenderich, B. (2009). Electric vehicles in the United States a forecasting model for hydrogen fuel cell vehicles considering infrastructure and
new model with forecasts to 2030. Berkeley: University of California. cost reduction effects. Energy Policy, 39(6), 3307e3315.
Boswijk, H. P., & Franses, P. H. (2005). On the econometrics of the bass diffusion Putsis, W. P., Jr. (1996). Temporal aggregation in diffusion models for first-time
model. Journal of Business and Economic Statistics, 23(3), 255e268. purchase: does choice of frequency matter? Technological Forecasting and So-
Cao, X. (2004). The future demand for alternative fuel passenger vehicles: A diffusion of cial Change, 51(3), 265e279.
innovation approach. Putsis, W., & Srinivasan, V. (2000). Estimation techniques for macro diffusion
Chandrasekaran, D., & Tellis, G. J. (2007). A critical review of marketing research on models. In E. Muller, Y. Wind, & V. Mahajan (Eds.), New-product diffusion models
diffusion of new products. N. K. Malhotra Review of Marketing Research, 3, (pp. 263e291). London: Kluwer Academic Publishers.
39e80. Schmittlein, D. C., & Mahajan, V. (1982). Maximum likelihood estimation for an
Cordill, A. (2012). Development of a diffusion model to study the greater PEV market (A innovation diffusion model of new product acceptance. Marketing, 1(winter),
Thesis Presented to The Graduate Faculty of The University of Akron In Partial 57e78.
Fulfillment of the Requirements for the Degree Master of Science). Schneider, M. (2002). Explorative Studie zur Markteinführung von Brenn-
Davidson, M., Cross-Call, D., Craig, M., & Bharatkumar, A. (2013). Assessing options €t Karlsruhe (TH).
stoffzellenfahrzeugen. Universita
for accommodating electric vehicles in Santa Delano Valley. USAEE Case Compe- Shoemaker, M. H. (2012). A bass diffusion model analysis: Understanding alternative
tition 2013. T. T. a. P. Group: 49. fuel vehicle sales. Claremont McKenna College.
Gross, U. (2008). Prognose des Absatzmarktes für alternative Antriebe e Modellbildung Shoemaker, M. H. (2012). A Bass diffusion model analysis: Understanding alterna-
und Simulation. Universita €t Karlsruhe (TH): 174. tive fuel vehicle sales. CMC senior thesis.
Heeler, R. M., & Hustad, T. P. (1980). Problems in predicting new product growth for Srinivasan, V., & Mason, C. H. (1986). Nonlinear least squares estimation of new
consumer durables. Management Science, 26(10), 1007e1020. product diffusion models. Marketing Science, 15(4), 169e178.
Hyman, M. R. (1988). The timeliness problem in the application of Bass-type new Steffens. (2003). A model of multiple-unit ownership as a diffusion process. Tech-
product growth models to durable sales forecasting. Journal of Business nological Forecasting and Social Change, 70(9), 901e917.
Research, 16(l), 31e47. Sultan, F., Farley, J. U., & Lehmann, D. R. (1990). A meta-analysis of applications of
Jensen, A. F., Cherchi, E., Mabit, S. L., & Ortúzar, J. d. D. (2014). Predicting the potential diffusion models. Journal of Marketing Research, 27(1), 70e77.
market for electric vehicles. Transportation Research Board. Takada, H., & Jain, D. (1991). Cross-national analysis of diffusion of consumer du-
Jiang, Z., Bass, F. M., & Bass, P. I. (2006). Virtual Bass model and the left-hand data- rable goods in Pacific rim countries. Journal of Marketing, 55(2), 48e54.
truncation bias in diffusion of innovation studies. International Journal of Van Den Bulte, C., & Stremersch, S. (2004). Social contagion and income hetero-
Research in Marketing, 23(1), 93e106. geneity in new product diffusion: a meta-analytic test. Marketing Science, 23(4),
Judge, G. G., Griffiths, W. E., Hill, R. C., Lütkepohl, H., & Lee, T.-C. (1986). The theory 530e544.
and practice of econometrics. Wiley, 1056. Van den Bulte, C., & Lilien, G. L. (1997a). Why Bass model estimates may be biased
Lamberson, P. J. (2008). The diffusion of hybrid electric vehicles. Future research (and what it means). The Pennsylvania State University, 7.
directions in sustainable mobility and accessibility. In Sustainable mobility Van den Bulte, C., & Lilien, G. L. (1997b). Bias and systematic change in the
accessibility research and transformation (SMART) at the University of Michigan parameter estimates of macro-level diffusion models. Marketing Science, 16(4),
center for advancing research and solutions for society (CARSS). 338e353.
Lee, D. H., Park, S. Y., Kim, J. W., & Lee, S. K. (2013). Analysis on the feedback effect Venkatesan, R., Krishnan, T. V., & Kumar, V. (2004). Evolutionary estimation of
for the diffusion of innovative technologies focusing on the green car. Techno- macro-level diffusion models using genetic algorithms: an alternative to
logical Forecasting and Social Change, 80(3), 498e509. nonlinear least squares. Marketing Science, 23(3), 451e464.
Li, G.. (nd). Power forecasting for plug-in electric vehicles with statistic simulations.