http://www.jstor.org/stable/184074 .
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a notforprofit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Marketing Science.
http://www.jstor.org
1.
Introduction
07322399/82/0101/0057$01.25
58
((P+^N^Ym
m N(t))))(m
(1)
)N(t
1).
59
[t 1,t]
E(X(t))
= (P +
N(t  1))(m
1)).
N(t
(2)
1) + a3N2(t
1) +
(t)
1) + E(t)
(3)
p= a/m
(4)
q= ma
(5)
and
(a
m=~~
2aj
 46163
*(6)
60
To follow up on this point, note that the left side of equation (2) should
theoretically be the derivative of N(t) and not the difference represented by
X(t). This substitution causes a problem in that, as defined, X(t) will underestimate dN(t)/dt for time intervals before the maximum adoption rate is
reached and will overestimate after that point.
The purpose of this paper is to propose and develop a maximum likelihood
approach to estimating the parameters of innovation diffusion models of new
product acceptance; more specifically, a new product growth model proposed
by Bass. It will be shown that the approach allows:
(i) computation of approximate standard errors for the diffusion model
parametersp, q, and m, and
(ii) determination of the required sample size for forecasting the adoption
level to any desired degree of accuracy. That sample size is shown to depend
both on the parameters determining the adoption process and the method
used to group the time series data.
The next two sections outline the maximum likelihood estimates (MLE) and
their sampling properties. These results are then used in ?4 to compare the
maximum likelihood formulation with OLS estimates above for data on eight
different product innovations. Finally, ?5 deals with the issue of required
sample sizes for parameter estimationparticularly in forecasting the ultimate
adoption level.
2. A Maximum Likelihood Approach
In his 1969 paper, Bass also states that, for eventual adopters, the likelihood
of purchase at time t given that no purchase has yet been made can be written
f(t)
f(t
1 F(t)
+ qF(t)
(7)
f(t)dt.
From equation (7) the probability density function for adoption at time t can
be expressed as:
f(t) = (p + qF(t))(1  F(t)).
Using the initial value condition that F(0) = 0, integration of the above
equation yields the cumulative distribution function (c.d.f.) for eventual
61
adopters:
(1  ebt)
F(t)
(1+
aeb
aebt)
where a = q/p and b = (p + q). Since this c.d.f. is clearly only appropriate
for eventual adopters the probabilities associated with it are conditional
probabilities. Thus in a system where the probability of eventually adopting is
c, the unconditional probabilities for adoption times are given by:
c(l  ebt)
b .
F(t) =
(8)
(1 + aeb)
dEt
E(N(t))
[cM
E(N(t))].
The fact that this is not equivalent to equation (1) illustrates an important
distinction between a model beginning with a distribution for adoption times
(equations (7)(8)) and one using the differential equation formulation (equation (1)(2)).2 Thus, while parameters p and q may still be interpreted as
coefficients of innovation and imitation (based on their impact on the c.d.f.and hence on expected behavior) they are not directly comparable to p and q
obtained by ordinary least squares.
Equation (8) represents the (c.d.f.) of adoption time for an individual
chosen at random from the population. In the applications of the model,
however, the individual adoption times are not known and instead a histogram with the number of individuals falling in each time interval is used to fit
the model. That is, let T equal the number of time intervals for which the data
are available and xi be the number of individuals who adopt the innovation in
time interval (ti_ , ti), i = 1,2, . . , T. Then typically to 0 and t  oo, with
Type I censoring occurring at time tT 1. In other words, one knows that XT
individuals did not adopt by time tT_ , but has no other information about
their adoption times, where by definition xT = M  Ci=l xi.
2We are grateful to Abba Krieger for pointing out this fact.
62
It has been well established (see, for example, Rao 1965) that under very
general regularity conditions, maximum likelihood estimates are best asymptotically normal. That is, they are consistent, asymptotically normal and
asymptotically efficient. In order to determine the maximum likelihood estimates of p, q, and m, we first generate the maximum likelihood estimates of a,
b, and c. Note that, due to the 11 correspondence between a, b, and c and p,
q, and m the maximum likelihood estimators for p, q, and m are easily
obtained from the maximum likelihood estimators for a, b, and c by the
following expressions:
p
C)(9)
(Ab
ab
(10)
(a + 1)()
and
mi= cM
(11)
Using equation (8), the likelihood function for the observed histogram of a
particular innovation can be written as:

L(a,b,c, xi) =[
TI
F(tIrl)
]T
i=1
[F(ti)
F(ti1)]xi
(12)
T1
l(a,b,c,xi)=
xi [lnc + In
i=l
+xln[c
Ile
bt,
bIbt
1+I ae  otI
,]
1+ ae  bt~_
ebt
+ aebtT

(13)
Note that in this model the impact of the diffusion process is felt through
the model's parameters. The coefficients of innovation and imitation determine the shape of the c.d.f., with such characteristics as a small value of q
implying slow expected growth, etc. Given these parameters, the likelihood
function (12) takes all individuals as independent draws from that particular
distribution (with given expected diffusion characteristics). Although more
direct word of mouth effects might be proposed, a tractable method for
incorporating them in this framework has yet to be developed.
63
v1
where the elements of the matrix V= [vij] are the expected values
v = E
(14)
aiasj
Uab
Obb
aac
bc
= V
(15)
Vab
Vbb
Vac
Vbc
Vcc
(16)
64
Using equations (13) and (14), the elements of the information matrix V,
equation (16), can be determined. The derivations of these elements are
included in the appendix. Knowing V, 2 could clearly be written out explicitly; but since the determinant and cofactors of V do not reduce appreciably
such a derivation did not seem worthwhile and was not undertaken. In an
actual application, it is easier to calculate V and then invert the 3 x 3 matrix
to find E. Note in the appendix that the covariance matrix, 2, depends on
both the true values of parameters a, b, and c, and the method of grouping
and censoring the adoption data (t,, . . , tT). (See equations A3A16.) Therefore, in addition to obtaining confidence limits for the parameters, the
efficiency of different data collection and report schemes can be evaluated.
Once the covariance matrix E for a, b, and c is obtained, the corresponding
variances and covariances of the MLE's pf,q, and m can be written using the
elements in equation (15) as: (see Kendall and Stuart, 1977, p. 247)
Obb
b2 a +
(
14
(a + 1)2
(a + 1a
a =
q = (a
bq2
1)+ (a
(a+
aq=
c(a
+ i)2b
a2
a + 1)
+
b(a  1)
a(ab' 18) (
(17)
(18)
()1)
b2
bca^
oa=
(a+,)2
b(ac
2b
2
ab
(a + 1)3
2ab
(a +
aac
auTc
(21)
(22)
M2bac
4.
EmpiricalResults
To illustrate the application of the maximum likelihood estimation procedure, time series data for four consumer durables and four medical technological innovations
were restricted to the early years of sales growth to help avoid replacement or
repeat purchases, and the data were collected from the Statistical Abstracts of
the United States.
65
All the published applications of the Bass model (including our application
to four consumer durables) have assumed the availability of the adoption time
series data on the entire population of the potential adopters. Since statistical
inference from the estimation equation (3) is not possible, the utilization of
such an application strategy is understandable. However, in some situations
such a data base may be unavailable or prohibitively expensive to construct.
As an alternative, one might estimate the life cycle of a new product from a
survey of the target market; For example, a company contemplating a new
product introduction may be interested in estimating the market potential (i.e.,
m) and industry growth rate (i.e., q). If the adoption census data are not
available, it may be necessary to estimate these parameters from a survey of
the target market consisting of both adopters and nonadopters.
With this in mind, survey data for four pieces of radiological equipment
(Ultrasound, CT HeadScanner, CT Body Scanner and Mammography) were
analyzed. The data were collected from 209 hospitals throughout the U.S.A.,
i.e., M = 209. The hospitals were asked to identify themselves as adopters/
nonadopters and, if adopters, provide the date of adoption (see Robertson and
Wind (1980) for details of this survey). The survey indicated that by 1978,
of the 209 hospitals, 168 (80.38%) had adopted Ultrasound, 113 hospitals
(54.07%) had adopted CT Head Scanner, 97 hospitals (46.41%) had adopted
CT Body Scanner and 119 hospitals (56.94%) had adopted Mammography.
4.1. ConsumerDurables
The four consumer durables examined were clothes dryers, room air conditioners, color televisions and dishwashers. In order to evaluate the effectiveness of maximum likelihood estimates, ordinary least squares estimates using
the differential equation model (equation (3)) were also obtained.
Table 1 summarizes the MLE and OLS estimates for the parameters p, q,
and m and also provides the estimated time of peak sales, t*. (For the sake of
brevity, regression coefficients a,, a2, and a3 in equation (3) are not reported.
Only the estimates of p, q, and m generated via equations (4)(6) are
provided). We stress again that although MLE and OLS estimates of p and q
have similar behavioral interpretations, they come from different models and
are not strictly comparable. The mean absolute deviation and mean squared
error are reported in Table 2, and the actual and fitted sales estimates for the
four products are depicted in Figures 14. Some important comments on these
results are warranted:
* For dishwashers, the OLS procedure yielded an inappropriate sign for the
regression coefficient a3, i.e., a, = 166.1017, a2 =.0985, a3 =.6874 x 105
(with R2 =.90). As equation (5) indicates, a positive sign for regression
coefficient (3 would give a negative sign for the parameter q. The MLE
estimates, however, possessed correct signs for all products analyzed. Naturally, the high R 2value supports the observation made by Heeler and Hustad
(1980) that the quadratic sales response function generally fits the data well.
* For clothes dryers, room air conditioners and color T.V. Figures 13 and
the statistics reported in Table 2 clearly indicate that the maximum likelihood
TABLE 1
Parameter Estimates for ConsumerDurables*
Product
Period
Covered
Clothes
Dryers
19491961
.0237
.3248
1949Room Air
Conditioners 1961
.0200
19631970
.0443
Color T.V.
Dishwashers
19491961
.3891
OLS
c
m(103)
.2937
.3249
15652
17314
t*
.0120
(.000003)
.3512
(.000087)
7.5114
7.2552
.0072
.4228
(.0000013) (.000074)
.6295
.5519
36117

3.9388
.0160
(.000002)
.6566
(.000080)
.0035
(.000005)
.1282
(.00018)
*For MLEs, numbers in parentheses are standard errors. Sample size taken as number
collection period:
# households, 1961: 53,291,000
# households, 1970: 65,444,000
67
TABLE 2
Fit Statistics for ConsumerDurables
Mean Absolute Deviation
MLE
OLS
Product
Clothes Dryers
Room Air
Conditioners
Color T.V.
Dishwashers
113.86
95.63
22286.14
16106.10
180.58
444.06

136.59
232.00
33.15
42354.81
317979.00

28759.00
121496.00
1688.04
estimates consistently provide a better fit to the data. In addition the standard
error generated by the maximum likelihood procedure indicate the general
stability of the estimated parameters.
In order to obtain some sense of the predictive validity of the two estimation approaches, onestepahead sales forecasts were generated for clothes
dryers, room air conditioners and color televisions. The fit statistics reported
in Table 3 again indicate the superiority of the maximum likelihood estimates
over those generated using the differential equation model with ordinary least
squares.
1500
1400
1300
1200
1100
1000
900
~
0f
~800
o?~?
t/
8?
//
_ __
0 ~C0
OLS Fitted
,
700
Actual
MLEFitted
600
500
400 E
/'
300
200
,
100
1949
Year
51
53
55
57
59
61
FIGURE1. Actual, MLE and OLS Fitted Number of Adopters for Clothes Dryers.
68
2100
1900
1700
S0
,_
1500
<
o
1300
1100
900
3 G_Q

700
Actual
MtE FITTED
OLS FITTED
500
Year
61
59
57
55
53
51
1949
FIGURE2. Actual, MLE and OLS Fitted Number of Adopters for Room Air Conditioners.
7000
6500
ur
L'i
w
ca
n
LA.
0
Q:
O
tY
LUi
co
._
j__
Actual
MLEFITTED
OLS FITTED
Year
1963
64
65
66
67
68
69
70
FIGURE3. Actual, MLE and OLS Fitted Number of Adopters for Color Televisions.
69
600
/,
500 
>O
Actual
O _
MLEFitted
/0
400 
X
o
200
100
FIGURE4.
4.2
53
51
1949
,,
55
Year
57
59
61
Medical Innovations
The maximum likelihood and ordinary least squares parameter estimates for
the four products analyzed are summarized in Table 4, with fit statistics given
in Table 5. The actual and fitted adoption estimates are plotted in Figures
58. Some important comments on these results follow:
* For Ultrasound and Mammography, the OLS procedure yielded a negative sign for the regression coefficient a&l.As equation (4) indicates, a negative
sign for regression a&would give a negative sign for either p or m.
TABLE 3
Product
Clothes Dryers
Room Air
Conditioners
Color T.V.
544.62*
589.29
3211.23
284.97
464299.00*
135904.00
426.47
2559.93
538381.00
11,664,233
245072.00
8,529,629
*The OLS procedure yielded inappropriate signs for the regression coefficients for the
period 19491955.
70
TABLE 4
Parameter Estimatesfor Medical TechnologicalInnovations*
Maximum Likelihood
q
c
p
Product
Ultrasound
CT Head Scanner
CT Body Scanner
Mammography
.00618
(.0025 )
.01252
(.00055)
.00988
(.00069)
.00256
(.00008)
.4342
(.0183)
.9875
(.0428)
.9901
(.0689)
.6462
(.0192)
.9377
(.0196)
.5886
(.0391)
.5894
(.0606)
.5942
(.0357)
OLS**
q
.06962
1.19165
.53110
.05719
1.78881
.42967
TABLE 5
Fit Statisticsfor Medical TechnologicalInnovations
Mean Absolute Deviation Mean Squared Error
OLS
MLE
OLS
MLE
Product
Ultrasound
CT Head Scanner
CT Body Scanner
Mammography
4.43
3.80
2.98
3.81
4.65
1.91
32.16
23.21

15.47
21.42
32.34
7.30
* Although the OLS procedure yielded correct parameter signs for Head
Scanner and Body Scanner, the estimates of the parameter c, the ultimate
penetration level, are less than the penetration levels already achieved by these
two products. As mentioned earlier, the survey identified 54.07% adopters of
CT Head Scanner and 46.41% adopters of CT Body Scanner. Both of these
adoption levels are higher than the ultimate penetration levels suggested by
the OLS procedure.
* The maximum likelihood estimates for all the four products possess
correct signs and their values are very plausible. The standard errors suggest
the general stability of these estimates. Furthermore, Table 5 and Figures 58
suggest that these estimates provide a good fit to the adoption data.
5.
71
28
25
Actual
0
Fitted
0  o
20
15
w
\\
0
cz
10
Year
1965
69
67
FIGURE 5.
71
73
75
79
77

Actual
O 0
Fitted
Year
1972
FIGURE 6.
73
74
75
76
77
78
72
DAVID C. SCHMITTLEIN
40 
Oo
Actual
0
Fitted
30
cc
20
/
/
/e
10 I
75
74
1973
76
'
Year
78
77
FIGURE7. Actual and MLE Fitted Number of Adopters for CT Body Scanners.
25
Actual
00
20 
 
Fitted
15
10
Year
1965
67
69
71
73
75
77
79
73
Ultrasound:
CT Head Scanner:
CT Body Scanner:
Mammography:
With these time intervals and the parameter values in Table 4, sample sizes
needed to forecast c to varying degrees of accuracy are given in Table 6. The
first two columns refer to 95% confidence intervals which are within 5 and 1
percentage points of the true market penetration. The last two columns
perform the same function for 99%confidence intervals.
Although no claims are made for generalizability of these numerical results,
the patterns which emerge in Table 6 are still noteworthy. The nearly tenfold
range in sample size needed across products indicates a great deal of sensitivi'y to the particular characteristics of the diffusion process being modeled.
Thus, the fact that a sample size of 1200 was sufficient for estimating market
penetration to within 5 percent must be taken with a note of caution.
Another interesting result concerns the difficulty of reducing the confidence
interval below 5 percent. Since the standard error declines proportionately
with vM, the number of observations must be increased by a factor of 25 to
narrow the confidence interval from ?5% to ? 1%. For this reason, an
alternative method of reducing the standard error may be considered in some
studies. Generally, lengthening the period of data collection (e.g., from 1978 to
1980) or collecting more frequent observations will also increase the reliability
of the parameter estimates. Repeating the calculations in ?3 for various data
collection schemes (i.e., choices of tl, .. ., t) one can compare the benefits of
each approach with an assessment of the expected costs.
TABLE 6
Sizes
RequiredSample
for Estimating Market Potential
Product
95%Confidence
? 01
?.05
99%Confidence
?.05
?.01
Ultrasound
CT Head Scanner
CT Body Scanner
Mammography
123
491
1179
409
213
847
2036
706
3084
12,275
29,485
10,233
5324
21,186
50,892
17,662
74
OLS
0.264
0.237
0.220
0.175
 0.42
MLE
0.188
0.122
 0.078
 0.089
 0.076
75
ab
Oac
Obb
= V
bc
(A1)
acc
V=
Vab
Vac
Vbb Vbc
(A2)
Vcc
with elements
T
Vaa
M[E
Vab
Vac
McBT[ 1+
2(a
i=1 Ai
(A3)
2Ci]
+ I)E  Di 
=
cM(a
Vbc=M(a+
],
(A4)
(A5)
Vbb= M(a + l)
1)
(a+ 1) A +
1))[ Dc
+
C
FJ
(A6)
(A7)
(A7)
T\
76
and
Vcc= M
(A8)
In equations (A3)(A8), Ai, Bi, Ci, Di, Ei, Fi, GT and HT are determined by
using the following equations expressed in terms of the estimated parameters
a, b, and c.
c 1  ebti
Ai = c 1
bt
+ ae bt1
Bi =
+ ae
)e
_ ebj
e2bti(
b _,
(AIO)
[1+
j
e2bti
c'
tie
[l+ae
bt,
(All)
+ aebt
[1
'e~bt*
D.=c C ,c
[l+ae'1
aebt',]
e2bt,(
'e
[1 + aebt"]
Ei =
el
bti
c
[1 + aebt]
EC=c
(A9)
']
,
i(A12)
[ + ae'1]
2bti
ti le 2bti_
bi]3
[ l+aebt"1]3'
(A13)
and
ti2et(l
 aebt)
[1 +aebt]3
c
ti2_lebti(
[1+
ae bti)
(A14)
aebt']3
The above equations (A9)(A14) can be used for i = 1,2, . .., T 1. For
i= T, however, the same equations can be used by setting the first term in
equation (A9) equal to one, and omitting the first term in equations (A10)
(A14) for
BT
77
Gr=
lc
tT lebt'(I
 e T
[1 + aeb']
bt
3)
(A15)
and
1 ebtTI
1 +
aebtT
(A16)
References
Bass, F. M. (1969), "A New Product Growth Model for Consumer Durables," Management
Science, 15 (January), 215227.
(1980), "The Relationship Between Diffusion Rates, Experience Curves, and Demand
Elasticities for Consumer Durable Technological Innovations," Journal of Business, 53,
S51S67.
Bretschneider, S. I. and V. Mahajan (1980), "Adaptive Technological Substitution Models,"
Technological Forecasting and Social Change, 18 (December), 129139.
Dodds, W. (1973), "An Application of the Bass Model in Long Term New Product Forecasting,"
Journal of Marketing Research, 10 (August), 308311.
Dodson, J. A. and E. Muller (1978), "Models of New Product Diffusion Through Advertising and
WordofMouth," Management Science, 24 (November), 15681578.
Fourt, L. A. and J. W. Woodlock (1960), "Early Prediction of Market Success for New Grocery
Products," Journal of Marketing, 25 (October), 3138.
Heeler, R. M. and T. P. Hustad (1980), "Problems in Predicting New Product Growth for
Durables," Management Science, 26 (October), 10071020.
Himmelblau, D. M. (1972), Applied Nonlinear Programming,McGrawHill, New York.
Horsky, D. and L. S. Simon (1978), "Advertising in a Model of New Product Diffusion,"
presented at the TIMS/ORSA National Meeting in New York City, May.
Kalwani, M. U. and A. Silk (1980), "Structure of Repeat Buying for New Packaged Goods,"
Journal of Marketing Research, 17 (August) 316322.
Kendall, M. G. and A. Stuart (1977), The Advanced Theory of Statistics, Vol. 1, Griffin, London.
Lilien, G. L., A. G. Rao and S. Kalish (1981), "Bayesian Estimation and Control of Detailing
Effort in a Repeat Purchase Diffusion Environment," Management Science 27 (May),
493506.
Mahajan, V. and E. Muller (1979), "Innovation Diffusion and New Product Growth Models in
Marketing," Journal of Marketing, 43 (Fall), 5568.
Mahajan, V. and R. A. Peterson (1978), "Innovation Diffusion in a Dynamic Potential Adopter
Population," Management Science, 15 (November), 15891597.
and (1979), "Integrating Time and Space in Technological Substitution Models,"
Technological Forecasting and Social Change, 14 (August), 231241.
Mansfield, E. (1961), "Technical Change and the Rate of Imitation," Econometrica, 29 (October),
741766.
Morrison, D. G., and D. C. Schmittlein (1980), "Jobs, Strikes and Wars: Probability Models for
Duration," OrganizationalBehavior and Human Performance,25, 224251.
Nevers, J. V. (1972), "Extensions of a New Product Growth Models," Sloan Management Review,
13 (Winter), 7791.
Peterson, R. A. and V. Mahajan (1978), "MultiProduct Growth Models," Research in Marketing,
Vol. 1, J. Sheth (editor), JAI Press, Greenwich, Conn., pp. 201232.
78
Rao, C. R. (1965), Linear Statistical Inference and Its Applications.Wiley, New York.
Robertson, T. S. and Y. Wind (1980), "Organizational Psychographics and Innovativeness,"
Journal of ConsumerResearch, 7 (June), 2431.
Robinson, V. and C. Lakhani (1975), "Dynamic Price Models for New Product Planning,"
Management Science, 21 (June), 11131132.
Rogers, E. M. and F. F. Shoemaker (1971), Communicationsof Innovations: A CrossCultural
Approach.The Free Press, New York.
Shenton, L. R. and K. O. Bowman (1977), Maximum Likelihood Estimation in Small Samples.
Macmillan, New York.
Taga, Y., and K. Isii (1959), "On a Stochastic Model Concerning the Pattern of Communications
Diffusion of News in a Social Group," Annals of the Institute of Statistical Mathematics, 11,
2543.
Wind, Y., V. Mahajan and R. C. Cardozo (1981), New Product Forecasting: Models and
Applications,Lexington Books, Lexington, Mass.
Lebih dari sekadar dokumen.
Temukan segala yang ditawarkan Scribd, termasuk buku dan buku audio dari penerbitpenerbit terkemuka.
Batalkan kapan saja.