combining related and sparse data in linear regression...
TRANSCRIPT
"COMBINING RELATED AND SPARSE DATA INLINEAR REGRESSION MCIDELS"
by
Wilfried VANHONACKERDonald LEHMANN**
andFareena SULTAN***
N° 89 / 19 (Revised March 89)
Associate Professor of Marketing, INSEAD, Boulevard de Constance,77305 Fontainebleau, France
George E. Warren Professor of Business, Graduate School of Business,Columbia University, New York, NY10027
Assistant Professor, Graduate School of Business Administration,Harvard University, Cambridge, MA02163
Director of Publication:
Charles WYPLOSZ, Associate Deanfor Research and Development
Printed at INSEAD,Fontainebleau, France
COMBINING RELATED AND SPARSE DATA IN LINEARREGRESSION MODELS
Wilfried R. Vanhonacker*Donald R. Lehmann**
andFareena Sultan***
Revised March 1989
* Associate Professor of Marketing, INSEAD, Boulevard de Constance,77305 Fontainebleau, France
** George E. Warren Professor of Business, Graduate School of Business,Columbia University, New York, NY10027
*** Assistant Professor, Graduate School of Business Administration, HarvardUniversity, Cambridge, MA02163
We would like to thank Gerry Tellis who was kind enough to share hisextensive data base on price elasticity estimates.
- 2 -
COMBINING RELATED AND SPARSE DATA IN LINEAR
REGRESSION MODELS
ABSTRACT
Meta-analysis has become a popular approach for studying
systematic variation in parameter estimates across studies. This
paper discusses the use of meta-analysis results as prior
information in a new study. While hierarchical prior distributions
in a traditional Bayesian framework are characterized by complete
exchangeability, meta-analysis priors explicitly incorporate
heterogeneity in prior vectors. This paper discusses the nature of
the meta-analysis priors, tFieir properties, and shows how they can
be integrated into a familiar recursive estimation framework to
enhance the efficiency of parameter estimates in linear regression
models. This approach has the added advantage that it can provide
such estimates when (a) the design or data matrix is not of full
rank, or (b) when observations are too few to allow independent
estimation. The methodology is illustrated using published and new
meta-analysis results in market response and diffusion of
innovation models.
Keywords: META-ANALYSIS; PRIOR INFORMATION; RECURSIVE ESTIMATION;
RANDOM COEFFICIENT REGRESSION; EXCHANGEABILITY.
- 3
1. INTRODUCTION
Meta-analysis has been promoted in a variety of fields as a way to
integrate research findings across various studies (Glass, McGaw, and Smith
1981, Hunter, Schmidt, and Jackson 1982, Peterson, Albaum, and Beltramini 1985,
Farley and Lehmann 1986, Hite and Fraser 1988, Sultan, Farley, and Lehmann
1988). Within an analysis-of-variance framework, the various studies are
viewed as imperfect replications of one overall but unplanned experiment. This
replication framework generalizes and enhances the understanding of systematic
variations in research findings based on design characteristics. Some
applications have appeared in the literature in recent years and a number of
interesting generalizations have been established and discussed (see, e.g.,
Assmus, Farley, and Lehmann 1984, Farley, Lehmann, and Ryan 1981, 1982,
Houston, Peter, and Sawyer 1983, Reilly and Conover 1983).
Meta-analysis findings, however, have rarely been incorporated as a source
of prior knowledge or information in subsequent studies. Given the design
characteristics of a new study, it should be rather simple to derive some prior
parameter estimates using meta-analysis results. When prior knowledge is
updated with sample information on the new situation, resulting parameter
estimates tend to be more stable and efficient (see, e.g., Lilien, Rao, and
Kalish 1981); hence, an attempt should be made to include prior information in
the estimation procedure.
Gain in efficiency from the use of prior information has been the reason
for extensive use and development of Bayesian statistical methods. The work by
Lindley and Smith (1972) which advocates the use of hierarchical prior distri-
butions has become the foundation for numerous applications. Early research on
integrating independently generated research findings into priors adopted this
approach (DuMouchel and Harris 1983, Kuczera 1983). However, the strong prior
assumption of exchangeability underlying hierarchical prior distributions has
been criticized as an oversimplification (Kass 1983). The assumption, usually
captured in a normal prior (Leaner 1978, p.49), states that the research
findings are related and that the data carry information about that relation-
ship. Because of differences in the research studies, heterogeneity is almost
certain to enter that relationship. Since prior information on heterogeneity
is often sparse, neither grouping nor the use of conditional priors has been
practical, and the traditional approach with its strict assumption of equal
information in all the prior studies has survived valid criticisms.
- 4 -
In attempting to uncover systematic variations, meta-analysis results
explicitly recognize parameter heterogeneity across independently executed
studies. Hence, using the results as prior information circumvents the strict
exchangeability assumption underlying hierarchical Bayes priors and addresses
the valid concern often raised in empirical research relying on such priors.
Using a known recursive estimator in a linear regression framework, this paper
discusses how these priors can be integrated with sample information. Besides
the established gain in parameter efficiency, the approach has the added
advantage that it can provide updated parameter estimates in instances where
sample information is missing or insufficient to allow for independent
estimation. The reported empirical illustrations are very encouraging in that
respect.
2. DERIVATION OF META-ANALYSIS PRIORS IN A MULTIPLE REGRESSION FRAMEWORK
Published meta-analysis focus on individual parameters analyzed
independently using ANOVA. As the parameters in each of these studies are not
estimated independently, this approach to recovering systematic variance in
them has ignored their covariance structures. Casting meta-analysis within a
multivariate regression framework allows the incorporation of those covariances
which, in a familiar Generalized Least Squares (GLS) framework, enhances the
efficiency of the results and sharpens the statistical power of the
significance tests. Although conceptually not different from an ANOVA
approach, the regression approach is adopted here as efficient priors are
naturally preferred over inefficient ones. Furthermore, this approach
simplifies comparisons to the traditional hierarchical Bayes method which it
improves upon by assuming partial exchangeability. To set the tone for
subsequent discussion and pinpoint inherent limitations of the suggested
approach, the meta-analysis priors are derived here within a multiple
regression framework.
Meta-analysis attempts to capture and explain systematic variations in
parameter estimates across various studies. Hence, a set of original studies
are available for which linear regression models can be specified as
y. = X i + u.
1 1 1for i 1, 2, ... m (1)
- 5
where y i is a (n x 1) vector containing observations on the dependent variable
in study i, X i is a (n x p) matrix containing observations on p non-stochastic
predictors in study i, Si is a (p x 1) vector containing the parameters, u i is
a (n x 1) disturbance vector, and m denotes the number of previous studies.
With respect to the disturbance terms, the assumptions are made that the
expected error is zero and that errors are independent both within and across
studies, or
E(u i ) = 0 for all i
2and E(u. 3 1= a. I for i j (2)1
0 otherwise.
In the current development of the methodology described hereafter, the
assumption of linearity in (1) is fundamental and represents a limitation.1
To the extent that prior studies cannot be cast in models which are linear in
parameters, the approach cannot be adopted. However, the methodology developed
subsequently is not confined to cases where disturbances are characterized by
scalar variance-covariance matrices as defined in (2). We confined ourselves
to this case here for simplicity.
The fundamental assumption of meta-analysis is that systematic variation
in the parameter vectors Oi can be related to study characteristics. For
example, systematic differences might be attributable to differences in product
category, industry, sample size, estimation method, etc. These study
characteristics define a multidimensional classification within which similar
studies can be grouped together. Within each cell, 1, of the classification,
the random coefficients assumption
0. + e.1 3
for all j (3)
holds with first and second moments
E(e.) = 0
for all j E k1 , 1 = 1, 2, ... k,
- 6 -
and E(e. = Ej 1
= 0
if i j and i, j"
otherwise.
In these expressions, k 1 denotes the set of studies in group 1 of k groups. In
hierarchical Bayes priors (see, e.g., Kuczera 1983), the parameter vectors Oi
for all i would be assumed to be drawn from a prior distribution (most likely
normal) with a mean vector and covariance matrix. Accordingly, the O i 's are
exchangeable and this prior distribution captures the "connectedness" between
all studies (DuMouchel and Harris 1983). In other words, each parameter vector
of a previous study contributes information about any single parameter in the
basic linear model. This strict assumption can be tested (see, e.g., Kuczera
1983), and lead to grouping of studies which could be analyzed separately.2
Capturing and describing the heterogeneity explicitly as achieved in meta-
analysis enables simultaneous analysis which enhances efficiency and
understanding.
The parameter vectors of the various studies can be combined as
0 = 0 + e
where 0' (13', R2 . ), P'' = ... ) and e' (el e' )
assuming studies similar in terms of design characteristics (i.e., partially
exchangeable) are grouped together. Note that vector a has m vector entries,
one for each original study. The vector entries corresponding to similar
studies are identical N vectors. In other words, if the first three studies
belong to group one, then /1 = /11 ...).
Of specific interest in meta-analysis are the variations in the mean
vectors al for 1 1, 2, Systematic variations arising from design
characteristics imply that a Zc and, hence,
S = Zc e (4)
- 7 -
That is, the mean parameter vectors can be expressed as a linear combination,
c, of specific design vectors which constitute the columns of matrix Z. Within
the multidimensional classification scheme adopted above, the design vectors in
matrix Z describe or explain group membership of the m original studies in the
k groups. These vectors could contain observations on continuous, categorical,
or dummy variables. Note that expression (4) not only describes the heteroge-
neity among the previous studies but at the same time captures the systematic
link among them.
In practice, the parameter vectors are not observed. Traditional meta-
analysis relies on parameter estimates in attempting to estimate vector c.
Hence, the basic model postulated is
0 = Zc + E (5)
whereW=(010',...0')with.containing unbiased least squares estimatesS.
of the parameters in study i, and E = e + u with u = 0 - 0 or the difference
between the estimated parameters and their true values (i.e., u i = (XiXi)Vu.
given (2)).
Given the assumptions about the random components in (2), (4), and (5), it
is evident that
E(E) = 0
2 _1E l + a l (X1X,) 0
E (EE') = Q =
2
(6)
0 E + a (X'X )k
m
Invoking Generalized Least Squares (GLS) principles given the non-scalar
variance-covariance matrix in (6), BLUE estimates of vector c in (4) can be
obtained which equal 3
and
- 8
_1 _1-c (Z'S? Z) Z'Q 0
(7)
Stated simply,
_1 _1 _1c DO with D (Z'Q Z) Z'Q (8)
In a new study, a regression model similar to the ones specified in (1)
can be postulated. Specifically, we have
yo = X 0 +u
oo o
2 4with E(u
o) = 0 and E(u u') = a
o I.
oGiven the meta-analysis results in (7),
a prior estimate of 00 can be obtained from
b = Z 00
where Zo defines the design of the current study in terms of the
characteristics incorporated in the design matrix Z of the meta-analysis.
As shown in Appendix 1, the prior estimate b can be expressed aso
b = Z DM(3, + Z Dd + ZoDy
o o o o (9)
where y = E + u with D as defined in (8), M being a matrix of m identity
matrices stacked on top of one another, and d being a vector containing the
difference between the mean parameter vector of the corresponding study and the
mean parameter vector for the studies belonging to the same set as the new
study. This is the meta-analysis prior which will be updated with sample
information in the new study using a familiar recursive estimator. The use of
this estimator fits nicely into the advocated regression approach and, as will
be shown shortly, makes the approach applicable when data in the new study are
either incomplete or insufficient for independent estimation.
- 9 -
3. RECURSIVE ESTIMATION
Updating the prior on 00 in (9) with sample information contained in yo
and Xo can be accomplished using the augmented system
Yo
=
Xo
uo
(10)
b - Z oDd
0 ZDM Z
oDy
When the prior information is with respect to a subset of the parameters in 00
(e.g.,°o1)'
the augmented
Yo
system
Xol
equals
Xo2
uo
°01 + 13o2
bol
- Zol
D1d1
Zol
D1M 0 Z
olD
1y1
where bol denotes the meta-analysis prior on 001 , and all other elements are
similar to the corresponding ones in (10) with 13,) = [g31 0,k] and X0 ,4X01 Xo2](where X
ol contains the columns in X
o corresponding to the parameters in S
ol ).
The system could be expanded further if independent prior information (e.g.,
separate meta-analysis results) was available for different subsets of 00.
This approach is similar to the Goldberger-Theil approach (Theil 1971)
with the apparent difference that (a) the prior values for 00 , b o , are adjusted
by ZoDd and (b) the matrix associated with vector 00 in the prior equation
(i.e. ZoDM in (10)) is different from an identity matrix. It is worth noting
that
b - Z Dd = Zo jc - Dd]0 o
- 10 -
and, hence,
_1
b - Z Dd = Zoc - Z
o(Z 1 4 Z) Z'9 d.
o o
Accordingly, Dd can be interpreted as a least squares estimate of the design
matrix Z on the differences in group means. If no groups were identified
(i.e., complete exchangeability), the prior would equal Z oc. Moreover, ZoDd is
subtracted from this prior to adjust for the partial exchangeability captured
in the profiles contained in design matrix Z. In essence, b o is adjusted in
the direction of the mean parameter vector of the group to which the new study
belongs.
Expressing (10) compactly, we have:
y = X00 + v (11)
where y' = [y' (b - Z Dd)']o o o
X' = [X'
and v' = [u' (2.0Dy)']•
With the variance-covariance matrix of v, g , as derived in Appendix 2 thev
recursive estimator of 0 equalso
1 _1 _1
0 = (X'Qv X) X'Q
v y
(12)
which has BLUE properties.
Some interesting observations can be made with respect to the estimator
shown in (12). The development thus far assumed that all information for the
new study was available. This is not always the case in practice. For
example, data on some predictors contained in X could be missing. Hence, the0
data matrix Xo would have columns containing zeros and a direct estimate of 0
o
could not be obtained.5 Note, however, that the augmented matrix X in (11) is
still of full rank; hence, a recursive estimate of all parameters in 0o is
obtainable despite the missing data. This is essentially an approach to data
transferability (Aigner and Learner 1984) which, using completely exchangeable
priors, has been studied in Vanhonacker and Price (1988).
Recursive estimates can also be obtained when the number of observations
in yo and X0 are smaller than the number of predictors specified in the basic
model (i.e., the number of columns in Xo). Again, direct least squares
estimates of 0o could not be obtained since X' X would not be of full rank.
o o
The augmented system in (10), however, does provide estimates for 0o in this
case with degrees of freedom equal to the number of observations actually
available on the dependent and independent variables.
4. SOME ESTIMATION ISSUES
2The above discussion assumed that E. (for i = 1, 2, ..., k), a. (for i =
0, 1, 2, ..., m), and d are known. In practice, these terms will have to be
2estimated. a
o can be estimated as the average of the variances for the
corresponding studies in the meta-analysis,
2 m, 2a = (l/m ) .E a
1 i= 1i
2wherea.=[(y. - X.0.)' (y. - X. .)( n - 1)) and 0i contains the parameter
estimates of the ith
study, and m1 denotes the number of previous studies
belonging to the first set (to which the new study belongs).
ThematricesE.for i 1, 2, ..., k can be estimated as the average of
the covariance matrices for the corresponding studies in the meta-analysis,
(13)
- 12 -
m.
.E1 (0. - 6.)(0. - (14)
m.
whereil.,..e1 (13./m.)withm.denoting the number of studies contained in the
.th1-- set. Note that assuming normality and invoking the Lindley and Smith
(1972) results, the estimates in (13) and (14) are related to the empirical
Bayes estimates using non-informative priors (see Aigner and Learner 1984).
2Estimatesfora.(i = 1, 2, ...m) can be obtained from the estimation
results of the previous studies. Specifically, the variance-covariance
matrices of the parameter estimates from the previous studies are estimates for
a. (X!X.)-1
(i = 1, 2, ...m) as incorporated in matrix 4, (see Appendix 2). The
difference between the mean parameter vectors contained in d can be derived
straightforwardly from the averaged estimates -6 i = j E l (0i /m i ) for i=1,2,...k.
5. EMPIRICAL ILLUSTRATIONS
Two different empirical illustrations are discussed. The first example
relies on two widely cited meta-analyses in the market response literature.
Both focus on a single but different parameter and were executed independently.
The second example relies on the meta-analysis of both parameters of the
linear, discrete analog of a popular model in the diffusion literature.
Market Response Models
The market response literature contains two meta-analyses: one on
advertising elasticities (Assmus, Farley, and Lehmann 1984) and one on price
elasticities (Tellis 1988). Accurate estimation of these parameters is
mandatory for optimal (e.g., profit maximizing) advertising spending and
pricing decisions. Elasticities are most often estimated using multiplicative
model specifications. Given that these can be linearized in parameters, the
procedure developed above can be implemented to enhance parameter efficiency
and stability in case of limited samples.
- 13-
In this illustration, time-series observations for a major brand in a low-
priced consumer product category were used together with the published meta-
analyses. Bimonthly observations were available on market share (the response
variable), advertising share, and relative price. Relative price was defined
as the brands price over the mean price of the three major competitors in the
category. These three competitors together account for over two-thirds of the
volume sold in each time period.
A multiplicative response model was specified as
In MSta + 0 In AS
t + y In RP
t + S In MS
t-1 u
t-1 t
where MSt denotes market share at time t, AS
t denotes advertising share at time
t, and RPt denotes relative price at time t. The parameters 0 and y denote,
respectively, the short-term advertising and price elasticities. Using the
meta-analysis prior on 0 from Assmus, Farley and Lehmann (1984) and the meta-
analysis prior on y from Tellis (1988), these parameters were estimated
recursively using five, six, seven, etc. sample observations. Before reporting
and discussing the results, a number of practical observations on the
implementation of the recursive estimator are in order.
Empirical studies seldom report standard errors of estimated parameters
(or exact t-values which would enable derivation of standard errors). The
data-base underlying Assmus, Farley, and Lehmann (1984) does not contain any
standard errors for the reported advertising elasticities. Tellis (1988)'s
data base contains some t-values but not for all price elasticities. Given
that published results are generally statistically significant, the standard
errors were derived assuming the corresponding t-values equaled three.6 With
this assumption, the entries for the matrices Q in (6) and T in (2-1) could be
derived. This assumption affected the weighting of the prior studies
incorporated in the meta-analysis data bases as well as the efficiency of the
derived parameter estimates.
Although the original data were used (with the original coding schemes),
the designs of both meta-analyses were reduced. This was done mainly to ensure
that the new study had a design profile which was matched in the data by at
least one study. Note that if that study belonged to an empty set, d in (9)
cannot be computed and, hence, no adjustment can be carried out as shown in
(10). The design reduction was done based on statistical significance. The
Assmus, Farley, and Lehmann (1984) design matrix Z was reduced from 26
Number of
Observations Meta-Analysis Prior on Singled
Parameter
Parameter Estimates (Estimated Variance)b
Based on
Sample observationsc Meta-Analysis Prior on Both
Parameters
Table 1
aELASTICITY ESTIMATES FOR MARKET SHARE RESPONSE MODEL
Advertising Price Advertising Price Advertising Price
5 0.2310(0.00) 1.7210(7.48) 0.0727(0.01) -2.3355(0.00) 0.0707(0.01) -2.3356(0.01)6 0.1530(0.01) -3.4815(12.8) 0.0633(0.00) -2.3363(0.00) 0.0639(0.00) -2.3359(0.01)7 0.1521(0.01) -3.4374(5.39) 0.0677(0.00) -2.3365(0.00) 0.0660(0.00) -2.3362(0.00)8 0.1544(0.00) -3.4914(4.67) 0.0684(0.00) -2.3365(0.00) 0.0672(0.00) -2.3363(0.00)9 0.1549(0.00) -4.0184(4.70) 0.0689(0.00) -2.3367(0.00) 0.0662(0.00) -2.3365(0.00)10 0.1520(0.00) -4.1684(4.05) 0.0679(0.00) -2.3368(0.00) 0.0650(0.00) -2.3366(0.00)11 0.0950(0.01) -3.5544(6.43) 0.0531(0.00) -2.3366(0.00) 0.0511(0.00) -2.3364(0.00)12 0.0462(0.01) -1.6317(10.7) 0.0372(0.00) -2.3357(0.01) 0.0383(0.00) -2.3356(0.01)13 0.0457(0.00) -3.5097(11.0) 0.0370(0.00) -2.3367(0.01) 0.0351(0.00) -2.3367(0.01)14 0.0554(0.01) -3.5970(9.87) 0.0408(0.00) -2.3367(0.01) 0.0390(0.00) -2.3367(0.01)15 0.0635(0.01) -4.2443(7.81) 0.0439(0.00) -2.3373(0.01) 0,0422(0.00) -2.3373(0.01)16 0.0582(0.01) -1.3461(5.05) 0.0420(0.00) -2.3348(0.01) 0.0431(0.00) -2.3348(0.01)17 0.0645(0.01) -1.0044(4.21) 0.0445(0.00) -2.3342(0.01) 0.0475(0.00) -2.3341(0.01)18 0.0548(6.00) -1.0553(3.78) 0.0452(0.00) -2.3342(0.01) 0.0558(0.00) -2.3340(0.01)19 0.0566(0.00) -1.0573(3.53) 0.0475(0.00) -2.3342(0.01) 0.0592(0.00) -2.3339(0.01)20 0.0610(0.00) -1.2296(3.26) 0.0504)0.00) -2.3344(0.00) 0.0605(0.00) -2.3341)0.00)
30 0.0739(0.00) -0.9645(1.45) 0.0640(0.00) -2.3320(0.00) 0.0642(0.00) -2.3320(0.00)
40 0.0937(0.00) -1.0894(0.99) 0.0806(0.00) -2.3294)0.01) 0.0832(0.00) -2.3293)0.01)
53 0.0493(0.00) 0.2452(0.55) 0.0473(0.00) -2.3083(0.01) 0.0563(0.00) -2.3112(0.01)
aMultiplicative response model with predictors advertising share, relative price, and lagged market share
bMeta-analysis priors derived from Assmus, Farley, and Lehmann (1984) and/or Tellis (1988)
cOLS estimates
dThe advertising elasticities shown were estimated using a prior on only that elasticity;the price elasticity estimates were estimated using a prior on only that elasticity
- 15 -
variables to the 11 most significant variables (including the intercept). The
Tellis (1988) design matrix Z was reduced from 21 variables to 10 variables
(including the intercept). The derived [bo - Z
o Dd] estimates equaled 0.0325
for the advertising elasticity and -2.3360 for the price elasticity.
It is important to note that the design profile of the new study matched
two studies out of 128 in the advertising-elasticity data base but only one out
of 421 in the price-elasticity data base. This limited representation has a
dual effect: first, the information contained in that subgroup will have
limited weight in the meta-analysis; second, as the adjustment of b o towards
the subgroup estimate is entirely a function of its members, if these are not
representative an inadequate adjustment can arise. These issues will have to
be kept in mind when evaluating the results, particularly for the price
response parameter.
The results are summarised in Table 1. Adding a single observation at a
time, the table shows the sample OLS estimates (no prior information), the
estimates using prior information on one of the two parameters, and the
estimates incorporating prior information on both parameters simultaneously.
The advertising parameter is very significant in the model and its estimate
gets adjusted very well in the direction of the "long-run" estimate (i.e., the
highly significant estimate over the entire sample observation period). For
example, the sample estimate with only five observations equals 0.2310;
incorporating prior information the estimate becomes, respectively, 0.0727 and
0.0707 which is much closer to the "long-run" estimate of 0.0493 and the
apparent equilibrium level between 0.05 and 0.06. For price, the sample
estimate is insignificant. Incorporating the corrected prior value of -2.336
draws the recursive estimate close to that value and makes it insensitive to
sample information. As the prior is adjusted towards a single study with a
highly significant price elasticity (i.e., a reported t-value of -8.33), the
significance of the recursive estimate is intuitively unreasonable given
limited representation. This raises an issue worth further investigation given
the otherwise very encouraging results.
New Product Diffusion: The Bass Model
In the modeling literature on the diffusion of innovations over time, the
Bass model has received considerable attention (Bass 1969). The basic premise
of the model can be stated as
- 16 -
P(t) = p + (q/m)Y(t) (15)
where P(t) denotes the probability of adoption at time t, Y(t) denotes the
cumulative number of adopters by time t, and p, q, and m are the three model
parameters.7 Accordingly, the probability of adoption at one point in time is
specified as a linear function in the number of adopters by that time. Hence,
the model describes the adoption process of an innovation as a contagion-type
process.
Following the algebra in the original paper by Bass (1969), it can be
shown that
2S(t) = a + bY(t) + c[Y(t)] (16)
where S(t) denotes the number of adopters at time t, and the parameters a, b,
and c are functions of the three original parameters as follows: a = pm,
b = q-p, and c = (-q/m). In discrete terms, model (16) is expressed as
St = a + bY
t-1 + cY
t-1 (17)
where t refers to a discrete time period (e.g., a month, quarter, year, etc.).
This quadratic equation was suggested by Bass (1969) as the basic model for
empirical analysis.
Since the original paper by Bass (1969), substantial research on the
specification and estimation of the model has been done. The specification has
been extended to incorporate explanatory variables such as price, advertising,
etc. (see, e.g., Mahajan and Peterson 1985). OLS estimation of the parameters
of model (17) has been shown to have a time aggregation bias (Srinivasan and
Mason 1986). Maximum likelihood estimation (Schmittlein and Mahajan 1982) as
well as Nonlinear Least Squares (NLS) of the continuous time Bass model
specification have been shown to be preferable over OLS with NLS being the
preferred estimator in terms of fit and predictive validity (Mahajan, Mason and
Srinivasan 1986, Srinivasan and Mason 1986). The linearity-in-parameters
constraint of the methodology suggested above confines us to the OLS estimation
of a, b, and c in the discrete analog model (17). Despite the shortcomings of
OLS, the estimates are easier to obtain and provide good starting values for
NLS. Within these constraints, a number of product/process innovations were
analyzed empirically.
- 17 -
A meta-analysis of Bass diffusion model estimates is contained in Sultan,
Farley, and Lehmann (1988). However, that study is very much in the tradition
of Assmus, Farley, and Lehmann (1984) and Tellis (1988) and is confined to an
independent analysis of individual parameters. In order to illustrate the full
power of the regression framework adopted above (which allows simultaneous
analysis of all parameters), a small meta-analysis was performed on the Bass
model results in Table 2.8
For simplicity of the illustration and in order to
preserve enough observations for the estimation of the variance-covariance
matrices in (14), only a single dummy predictor was specified in (5). In other
words, the various innovations listed in Table 2 were broken down into 2
groups. The variable was whether or not the innovation was a consumer durable
product innovation. A distinction was made between consumer and industrial
durables because clear differences exist in their diffusion process
characteristics. The absolute magnitudes of the potential adopter populations
are very different with consumer durables generally having much larger
population figures than non-consumer durables. This results in striking
differences in the values of, for example, the parameter "a" in model (17)
which captures the number of adopters in the initial period of market
introduction.
As indicated in the derivation of expression (7), the covariances of the
estimated parameters should be incorporated in order to secure efficient
estimation of the meta-analysis parameters (i.e., vector c in (5)). However,
these covariances are not reported in the studies whose results are shown in
Table 2. Only the variances or standard deviations of the parameter estimates
are given. Accordingly, vector c in expression (5) was estimated with
variance-covariance matrices in which all off-diagonal elements were zero.
The intercept estimates for parameters a, b, and c (representing the
coefficients for an industrial product) were equal to, respectively, 8.52905,
0.56810, and -0.0049512; the additive effect for a consumer durable innovation
equaled, respectively, 641,196.0, -0.25537, and 0.0049511. These additive
effects essentially indicate that (as discussed above) the potential adopter
populations for consumer durables are much larger than those for industrial
innovations, and consumer durable innovations diffuse slower on their own than
industrial innovations. The R equaled 0.36. Given cross-sectional observa-
tions and a single predictor, this fit is impressive.
The methodology incorporating the meta-analysis results discussed above
was applied to forecasting sales of color televisions. The sales data were
obtained from Wind and Mahajan (1985, p.215) and consist of annu 1 observations
Table 2
Bass Model Regression Results for Consumer/Industrial Innovation
2
PRODUCT/TECHNOLOGY Time 11 of ob-
Period servations
3 -7
Electric Refrigerators '20-'40 21 104.67(10 ) 0.21305 -0.053913(10 ) 0.903
3 -7Home Freezers '46-'61 16 308.12(10 ) 0.15298 -0.077868(10 ) 0.742
3 -7
Black and White TV's '46-'61 16 2,696.20(10 1 0.23317 -0.025957(10 1 0.576
3 -7
Room Airconditioners '46-'61 16 175.69(10 ) 0.40820 -0.247770(10 ) 0.911
3 -7
Clothers Dryers '48-'61 14 259.67(10 ) 0.33968 -0.236470(10 ) 0.896
3 -7Power Lawnmowers '48-'61 14 410.98(10 ) 0.32871 -0.075506(10 ) 0.932
3 -7
Electr. Bed Coverings '49-'61 13 450.04(10 ) 0.23800 -0.031842110 ) 0.976
3 -7
Automatic Coffee Makers '48-'61 14 1,008.20(10 ) 0.28435 -0.051242(10 1 0.883
Steam Irons '49-'60 12 1,594.70(103) 0.29928 -0.058875(1 -7) 0.828
3 -7
Record Players '52-'61 10 543.94(10 ) 0.62931 -0.298170(10 ) 0.899
Holiday Inn Motels '54-'68 15 9.27866 0.35521 -0.000258 0.918
Howard Johnson '56-'69 11 10.64810 0.23687 -0.000502 0.759
Motor Lodges
Ramada Inns '62-'69 8 14.29150 0.14904 0.004850 0.746
Howard Johnson, '56-'69 11 24.58870 0.32699 -0.000150 0.930
Holiday Inn,
Ramada Inns
McDonald's Restaurants '55-'65 11 15.00440 0.52016 -0.000646 0.964
Rapid Bleach Proc. '59-'68 10 1.86071 0.54948 -0.014050 0.919
Conversion to 70% HO2 2
'62-'68 7 5.26971 1.02542 -0.006090 0.974
Deliv. Syst.
Continuous Bleach '44-'51 8 2.63906 0.90471 -0.014600 0.827
Ranges
Hybrid Corn '13-'41 9 3.51145 0.96654 -0.011107 0.926
aPublished results from Bass (1969) end Nevers (1972)
- 19 -
from 1963 until 1970.Some observations are in order with respect to the product
and the data. The Bass model has been fitted quite successfully to the color
television data (Bass 1969, Nevers 1972). Predictions of peak sales (which
occur within the period of observation) and the magnitude of the peak have been
quite accurate. Comparison of the results derived here with the published
independent Bass model predictions will constitute a relatively strong test of
the power of the recursive estimator.
The recursive estimation results and their implied estimates for the
parameters of model (15) are shown in Table 3.9 Three sets of results are
shown: one set of prior estimates derived from the meta-analysis estimates, one
set derived from the priors and the first annual sales observations for color
televisions, and one set derived from the priors and the first two annual sales
observations. Estimation results with more data points converged to the least
squares estimates based exclusively on color television sales observations
published in Bass (1969) and Nevers (1972). This rapid convergence indicates
that the sample observations contain much more information than the prior
information. This is intuitively reasonable since color televisions, relative
to the innovations listed in Table 2, are more an improvement of an existing
product (i.e., black/white television) than a discontinous innovation.
Accordingly, one might expect the faster rate of diffusion observed in the
recursive estimates in Table 3 as sample observations are added.
The corresponding unit sales forecasts are shown in Table 4. The pattern
of rapid diffusion is evident when comparing actual sales to the predictions
based on the prior estimates. Adding the first year sales figure adjusts the
forecasts in the right direction but there is still a significant
underprediction of actual sales. When the second year is added, the pattern is
adjusted further. A reasonable forecast is obtained with a slight
overprediction in the post sample years. This early 1965 forecast predicts
sales to peak in 1967 below the 7 million unit point. This forecast is more
acurate and much more reasonable than the industry expectations at that time
10(and even later; see Bass 1969, p.224). With a better-fitting meta-
analysis model than the one relied on in this illustration, the early forecast
is likely to improve further.
- 20-
Table 3
Estimation Results for Color Televisions
Parameter
Recursive Estimation
Meta-Analysis
Prior
Number of Observations Incorporated
1 2
a(103 )
b
7c(10
m(106
)
P
q
T*a
S(T*)(106
) b
)
641.205
0.31273
-0.10810
30.852
0.021
0.334
7.834
2.903
747.000
0.41389
-0.13918
31.445
0.024
0.438
6.315
3.824
747.000
1.01309
-0.42612
24.489
0.031
1.044
3.289
6.579
a Time of peak sales, or T* . [1/(p+q)7[1n(q/p)] (Bass 1969).
bMagnitude of peak sales, or S(T*) = [m(p+q) 2
]/4q (Bass 1969).
- 21 -
Table 4
Forecasts for Color Televisions
Forecast based on
Recursive Estimation
Number of Observations
Meta-Analysis
Incorporated
Actual
Prior
1 2
Time 1968 1972 1970 1967
Peak Sales
Magnitude 5.982 2.903 3.824 6.579
(106
)
Sales (106
) 1963 0.747 0.641 0.747 0.747
1964 1.480 0.837 1.048 1.480
1965 2.646 1.077 1.445 2.792
1966 5.118 1.370 1.942 4.758
1967 5.777 1.702 2.518 6.579
1968 5.982 2.059 3.109 5.915
-22-
6. SUMMARY AND CONCLUSION
Meta-analysis has been postulated in a variety of fields as a way to
integrate research findings across studies. Within an analysis-of-variance
framework, the parameter estimates of the various studies are viewed as
imperfect replications of one overall but unplanned experiment. This
replication framework generalizes and enhances the understanding of systematic
variations in research findings.
Meta-analysis results have never been used, however, as a source of prior
information in subsequent studies. Within the design underlying the meta-
analysis, it is rather straightforward to obtain prior estimates for the
parameters of a new study. This paper described a methodology to update this
prior information with sample information for the new study. The methodology
relies on a random coefficients regression framework and accomplishes the
updating task using a recursive estimator. The procedure is versatile and
allows for efficient estimation even when independent estimation is infeasible
because the data are either incomplete or too sparse. Empirical examples in
the market response modeling literature and in the diffusion modeling
literature are encouraging.
-23 -
FOOTNOTES
1. Linearity is an implicit assumption in ANOVA analyses.
2. Prior grouping is, of course, also possible but general lack of priorinformation on group membership makes this impractical.
3. The result shown in (7) suggests that the most efficient estimator of c inmodel (5) is the GLS estimator shown in (7).
4. Whether or not the similarity in model specification is justified dependson the problem studied and is largely determined by theoretical issues.Note, however, that if there is little similarity or analogy betweenprevious studies (incorporated in the meta-analysis) and the new study,the meta-analysis prior will be a non-informative prior.
5. Replacing missing values with zeros is the traditional approach in datatransferability (see e.g., Aigner and Learner 1984). Other approaches havebeen suggested in the literature on missing data (see, e.g., Dunsmuir andRobinson 1981, Gourieroux and Monfort 1981, Steward and Sorenson 1981).These approaches rely on likelihood functions and are often impractical.Working with zeros enables recursive estimation as a simple, practical,and conceptually appealing method to the problem investigated.
6. The importance of incorporating the estimated standard errors isemphasized in Montgomery and Srinivisan (1989).
7. At the aggregate level, parameter p in (15) denotes that fraction of thepopulation who adopt at a particular time whose decision to adopt was notinfluenced by the social environment (hence, the labeling of p as thecoefficient of "innovation"). Parameter q in (15) is labeled thecoefficient of "imitation" as it captures the effect of units sold on thelikelihood of adoption. Parameter m in (15) denotes the size of thepopulation which will ultimately adopt the innovation.
8. Water softeners and boat trailers from, respectively, Bass (1969) andNevers (1972) were not incorporated because of unrealistic (perhapsmisprinted) values.
2 29. Since a. for i = 1,2,...,m were not available for previous studies, a
o as
2shown in (13) could not be derived. The value for a
o incorporated in the
variance-covariance matrix 4 was set equal to the unit sales figure inv
year one.
10. With more than two sample observations, the results (not shown) converge tothe accurate forecasts shown in Bass (1969) and Nevers (1972).
- 24 -
Appendix 1: Derivation of Prior Estimate
Given expression (8), the prior estimate equals
b = Z DOo o
and, since 0 = 0 + u with terms as specified above,
b Z DO + ZoDu.
o o
Depending on the design of the new study, it will belong to one of the k
groups specified in terms of design characteristics in the meta-analysis. If
we arbitrarily assume that the study belongs to the first group (k 1 ), we can
write the parameter vector as
i3o = 1
eo
.
Hence,
0. = 0 + e. for all j k, (1-2)
o io
wheree.=e.-eo. For any other study 1 not belonging to set k 1 , we have
o
- 130(6i - 61) elo (1-3)
with 0i denoting the mean parameter vector for the set of studies containing
study 1.
Combining expressions (1-2) and (1-3) for all studies, we can specify the
augmented model
0 = M00 + d +
where E = [e'[eo e'
oe' I, and M denotes a matrix of m identify matrices1 2 MO
stacked on top of one another. Vector d contains zeros in the entries
corresponding to the parameter vectors for all studies belonging to the same
set as the new study, and the difference in mean parameter vectors (i.e.,
1) in the entries corresponding to parameters vectors for all studies
61
which do not belong to the same set as the new study. Accordingly, the prior
estimate bo in (1-1) can be expressed as shown in (9).
- 25-
Appendix 2: Computation of Variance-Covariance Matrix Qv.
Given expression (11)
21
a 0
E(vv') = 4v=
0 ZoDE(yy')D'Z'
o
21
a 0
0 ZoDyD'Z'
o
where E(yy') = y. Accordingly, y = ER + u)(E. + u)']
Assuming that E and u are independent and knowing that the new study belongs to
the first set, it can be shown rather easily that
22E 1+ a,(X1X,)
El
E,
E1
2E,
E,
2+ a 2 (X'2 X 2 )
E1
E1
E,
E,
E,
(2-1)
2
(E i +Ek) + a (X'X )-1
M M M
or simply y + E6) E,] where E denotes an (m x m) matrix containing all ones
and C)denotes a Kronecker product. Note that from the expression of the first
two diagonal elements in Y, it is evident that the corresponding prior studies
belong to the set which also contains the new study.
El E,
- 26 -
REFERENCES
Aigner, D.J. and E.E. Learner (1984), "Estimation of Time-of-Use PricingResponse in the Absence of Experimental Data," Journal of Econometrics,26, 205-227.
Assmus, G., J.U. Farley, and D.R. Lehmann (1984), "How Advertising AffectsSales: Meta-Analysis of Econometric Results," Journal of MarketingResearch, 21, 65-74.
Bass, F.M. (1969), "A New Product Growth Model for Consumer Durables,"Management Science, 15, 215-227.
DuMouchel, W.H. and J.E. Harris (1983), "Hayes Methods for Combining Resultsof Cancer Studies in Humans and Other Species," Journal of the AmericanStatistical Association, 78, 293-308.
Dunsmuir, W. and P.M. Robinson (1981), "Estimation of Time Series Models in thePresence of Missing Data," Journal of the American StatisticalAssociation, 76, 560-568.
Farley, J.U., D.R. Lehmann and M.J. Ryan (1981), "Generalizing from 'Imperfect'Replication," Journal of Business, 54, 597-610.
Farley, J.U., D.R. Lehmann and M.J. Ryan (1982), "Patterns in Parameters ofBuyer Behavior Models: Generalizing from Sparse Replication," MarketingScience, 1, 181-204.
Glass, G.V., B. McGaw and M.L. Smith (1981), Meta-Analysis in Social Research,Beverly Hills, CA: Sage Publications, Inc.
Gourieroux, C. and A. Monfort • (1981), "On the Problem of Missing Data in LinearModels," Review of Economic Studies, 48, 579-586.
Heeler, R.M. and T.P. Hustad (1980), "Problems in Predicting New Product Growthfor Consumer Durables," Management Science, 26, 1007-1020.
Hite, R.E. and C. Fraser (1988), "Meta-Analysis of Attitudes Toward Advertisingby Professionals," Journal of Marketing, 52, 95-103.
Houston, M.J., J.P. Peter and A.G. Sawyer (1983), "The Role of Meta-Analysis inConsumer Behavior Research," in Advances in Consumer Research, Vol. 10,R.P. Bagozzi and A.M. Tybout (eds.). Ann Arbor, MI: Association forConsumer Research.
Hunter, J.E., F.L. Schmidt and G.B. Jackson (1982), Meta-Analysis: CumulatingResearch Findings Across Studies. Beverly Hills, CA: Sage Publications,Inc.
Kass, R.E. (1983), "Comment," Journal of American Statistical Association, 78,312-313.
Kuczera, G. (1983), "Improved Parameter Inference in Catchment Models 2.Combining Different Kinds of Hydrolic Data and Testing TheirCompatibility," Water Resources Researc', 19, 1163-1172.
- 27 -
Learner, E.E. (1978), Specification Searches. New York, N.Y.: J. Wiley and Sons.
Lilien, G.L., A.G. Rao and S. Kalish (1981), "Bayesian Estimation and Controlof Detailing Effort in a Repeat Purchase Diffusion Environment,"Management Science, 27, 493-506.
Lindley, D.V. and A.F.M. Smith (1972), "Bayes Estimates for the Linear Model,"Journal of the Royal Statistical Society, B34, 1-41.
Montgomery, D.B. and V. Srinivisan (1989), "An Improved Method for MetaAnalysis: With Application to New Product Diffusion Models," presented atthe ORSA/TIMS 1989 Marketing Science Conference, March 15-18, DukeUniversity.
Nevers, J.V. (1972), "Extensions of a New Product Growth Model," SloanManagement Review, 13, 77-91.
Peterson, R.A., G. Albaum and R.P. Beltramini (1985), "A Meta-Analysis ofEffect Sizes in Consumer Behavior Experiments," Journal of ConsumerResearch, 12, 97-103.
Reilly, M.D. and J.D. Conover (1983), "Meta-Analysis: Integrating Results fromConsumer Research Studies," in Advances in Consumer Research, Vol.10, R.P.
Bagozzi and A.M. Tybout (eds.). Ann Arbor, MI: Association for ConsumerResearch.
Schmittlein, D.C. and V. Mahajan (1982), "Maximum Likelihood Estimation for anInnovation Diffusion Model of New Product Acceptance," Marketing Science,1, 57-78.
Srinivasan, V. and C.H. Mason (1986), "Nonlinear Least Squares Estimation ofNew Product Diffusion Models," Marketing Science, 5, 169-178.
Stewart, W.E. and J.P. Sorenson (1981), "Bayesian Estimation of CommonParameters from Multiresponse Data with Missing Observations,"Technometrics, 23, 131-146.
Sultan, F., J.U. Farley and D.R. Lehmann (1988), "A Meta-Analysis of DiffusionModels," Working Paper, Columbia University.
Tellis, G.J. (1988), "The Price Elasticity of Selective Demand: A Meta-Analysis of Econometric Models of Sales," Journal of Marketing Research,25, 331-341.
Theil, H. (1971), Principles of Econometrics. New York, NY: J. Wiley
Tigert, D. and B. Farivar (1981), "The Bass New Product Growth Model: ASensitivity Analysis for a High Technology Product," Journal of Marketing,45, 81-90.
Vanhonacker, W.R. and L.J. Price (1988), "Data Transferability: EstimatingThe Response Effect of Future Events Based on Historical Analogy," WorkingPaper, No.88/57, Insead, Fontainebleau, France.
Wind, Y. and V. Mahajan (1985), eds., Innovation Models of New Product Acceptance, Boston, MA: Ballinger Publishing.
INSRAD WORK/NC PAPERS SERIES
"The R i D/Production interface".
"Subjective estimation in integrating
communication budget and allocationdecisions! • case study', January 1986.
• Sponsorship and the diffusion oforganisational innovation: a preliminary viev'.
"Confidence intervals: an empiricalinvestigation for the series in the N-
Competition" .
• note on the reduction of the workweek",
July 1985.
"The real exchange rate and the fiscalaspects of • natural resource discovery",
Revised version: February 1986.
'Judgmental biases in sales forecasting',
February 1986.
"Forecasting political risks forinternational operations", Second Draft:
March 3, 1986.
"Conceptualizing the strategic process indiversified firma: the role and nature of thecorporate Influence process • , February 1986.
"Analysing the issues concerning
technological de-maturity'.
"From 'Lydiametry" to "Pinkhamization":
misspecifying advertising dynamics rarely
affects profitability'.
"The economics of retail firms', RevisedApril 1986.
'Spatial competition it la Cournot'.
nomparaison internetionale des merges brutes
du commerce', June 1985.
'Les prime■ des offres publiques, le noted'informalion et le marche des transferts deconical* des societies'.
'Strategic capability transfer In acquisition
integration", May 1986.
"Towards an operational definition of
services', 1986.
"Nostradsaus: • knovledge-based forecasting
advisor•.
'The pricing of equity on the London stockexchange: seasonality and size premium',
June 1986.
"Risk-pttais seasonality in U.S. and European
equity markets", February 1986.
'Seasonality in the risk-return relationships
some international evidence', July 1986.
•An exploratory study on the integration ofinformation systems in manufacturing',
July 1986.
"A methodology for specification andaggregation in product concept testing',
July 1986.
•Protection", August 1986.
'The economic consequences of the FrancPoincare', September 1986.
'Negative risk-return relationships inbusiness strategy: paradox or truis■?",
October 1986.
'Interpreting organizational texts.
'Why follov the leader?".
•The succession gam: the real story.
"Flexibility: the next competitive battle',
October 1986.
"Flexibility: the next competitive battle',
Revised Version: March 1987
1986
86/01
Arnoud DE MEYER
86/02 Philippe A. NAERT
Marcel WEVERBERG8and Guido vERSLIIJVEL
86/03 Michael BRIER
86/04 Spyros MAKRIDAKISand Michele RIBON
86/05 Charles A. VYPLOSZ
86/06 Francesco GIAVA22I,Jeff R. SHEN and
Charles A. VYPLOSZ
86/07 Douglas L. MacLACHLANand Spyros MAKRIDAKIS
86/08 Jost de le TORRE and
David R. NECKAR
86/09 Philippe C. RASPESLACH
86/10 R. MOENART,
Arnoud DE MEYER,
J. DAME and
D. DESCROOLMEESTER.
86/11 Philippe A. NAERT
and Alain BULTE2
86/12 Roger BETANCOURTand David CAUTSCHI
86/13 S.P. ANDERSON
and Damien J. NEVEN
86/14 Charles WALDMAN
86/16 B. Espen ECKBO andHervig M. LANGOMR
86/17 David B. JEMISON
86/18 James TEBOULand V. MALLERET
86/19 Rob R. vniTz
86/20 Albert CORRAY,Gabriel HAwAvINIand Pierre A. ?gun
86/21 Albert CoRHAY,Gabriel A. RAVAVINIand Pierre A. MICHEL
86/22 Albert CORRAY,
Gabriel A. RAVAWINI
and Pierre A. MICHEL
86/23 Arnoud DE MEYER
86/24 David CAUTSCRI
and Vithala R. RAO
86/25 R. Peter CRAY
and Ingo WALTER
86/26 Barry EICRENGREENand Charles VYPLOSZ
86/27 Karel COOL
and Ingemar DIERICKX
86/28 Manfred KETS DE
VRIES and Danny MILLER
86/29 Manfred KETS DE VRIES
86/30 Manfred KETS DE VRIES
86/31 Arnoud DE METER
86/31 Arnoud DE METER,
Jiniehiro NAKANE,
Jeffrey C. MILLERand Kasra FERDOVS
86/15 Mihkel TOMBAK end
Arnoud De METER
"Boy the managerial attitudes of firas vith
FMS differ from other manufacturing floss:
surre y results'. June 1986.
86/32 Karel COOLand Dan SCHENDEL
Perforasnee differences saong strategic group
wembere, October 1986,
86/33 Ernst BALTENSPERGERand Jean DERM/NE
86/34 Philippe FIASPESLAGHand David JEMISON
86/35 Jean DERMINE
86/36 Albert CORHAT andGabriel HAWAWINI
86/37 David GAUTSCIII andRoger BETANCOURT
86/38 Gabriel BAVAVINI
86/39 Gabriel NAVAVINIPierre MICHELand Albert COREA/
86/40 Charles VYPLOSZ
86/41 Kasra FERDOVSand Wickham SKINNER
86/42 Kasra FERDOWSand Per LINDBERG
86/43 Damien NEVEN
86/44 Ingemar DIERICKXCarmen MATUTESand Damien NEVEN
1987
97/01 Manfred KETS DE VRIES
87/02 Claude VIALLET
87/03 David GAUTSCRIand Vithala RAO
87/04 Sumantra CHOSRAL andChristopher BARTLETT
87/05 Arnoud DE MEYERand Rasra PERDOVS
"The role of public policy in insuringfinancial stability: a cross-country,comparative perspective", August 1986, RevisedNovember 1986.
"Acquisitions: myths and reality',July 1986.
"Measuring the market value of a bank, aprimer', November 1986.
' Seasonality in the risk-return relationship:some international evidence', July 1986.
"The evolution of retailing: • suggestedeconomic interpretation'.
'Financial innovation and recent developmentsin the French capital markets*, Updated:September 1986.
'The pricing of common stocks on the Brusselsstock exchanger a re-exaxination of theevidence', November 1986.
' Capital flows liberalization and the EMS, aFrench perspective', December 1986.
'Manufacturing in a new perspective',July 1986.
' FMS as indicator of manufacturing strategy',December 1986.
'On the existence of equilibrium in hotelling'amodel', November 1986.
' Value added tax and competition',December 1986.
'Prisoners of leadership'.
'An empirical investigation of internationalasset pricing', November 1986.
' A methodology for specification andaggregation in product concept testing',Revised Version: January 1987.
'Organizing for innovations: case of themultinational corporation', February 1987.
' Managerial focal points in manufacturingstrategy', February 1987.
'Customer loyalty as a construct in themarketing of banking services", July 1986.
"Equity pricing and stock market anomalies',February 1987.
' Leaders vho can't manage', February 1987.
"Entrepreneurial activities of European MBAs",March 1987.
'A cultural view of organizational change',March 1987
"Forecasting and loss functions", March 1987.
' The Janus Bead: learning from the superiorand subordinate faces of the 'manager's job',April 1987.
"Multinational corporations as differentiatednetworks", April 1987.
'Product Standards and Competitive Strategy: AnAnalysis of the Principles', May 1987.
' KETAFORECASTINC: Ways of improvingForecasting. Accuracy and Usefulness",May 1987.
' Takeover attempts: what does the language tellus?, June 1987.
"Managers' cognitive maps for upward anddovnvard relationships', June 1987.
"Patents and the European biotechnology lag: astudy of large European pharmaceutical firms',June 1987.
"lay the EMS? Dynamic games and the equilibriumpolicy regime, May 1987.
'A new approach to statistical forecasting",
June 1987.
"Strategy formulation: the impact of nationalculture', Revised: July 1987.
' Conflicting ideologies: structural andmotivational consequences', August 1987.
'The demand for retail products and thehousehold production model: new views oncoaplementarity and substitutability*.
07/06 Arun K. JAIN,Christian PINSON andNaresh K. MALMOTKA
87/07 Rolf BANE andGabriel HAVAVINI
87/08 Manfred KETS DE VRIES
87/09 Lister VICKERY,Mark PILKINGTONand Paul READ
87/10 Andre LAURENT
87/11 Robert FILDES andSpyros MAKRIDAKIS
87/12 Fernando BARTOLOMEand Andre LAURENT
87/13 Sumantra GHOSHALand Nitin NOIIRIA
87/14 Landis GABEL
87/15 Spyros HAKRIDAKIS
87/16 Susan SCHNEIDERand Roger DUNBAR
87/17 Andre LAURENT andFernando BARTOLOME
87/18 Reinhard ANCELMAR andChristoph LIEBSCHER
87/19 David BEGG andCharles WYPLOS2
07/20 Spyros MAKRIDAKIS
87/21 Susan SCHNEIDER
87/22 Susan SCHNEIDER
87/23 Roger BETANCOURTDavid GAUTSCHI
87/24 C.B. DERR andAndre LAURENT
87/25 A. K. JAIN,N. K. MALHOTRA andChristian PINSON
87/26 Roger BETANCOURTand David GAUTSCHI
07/27 Michael BURDA
87/28 Gabriel HAVAVINI
87/29 Susan SCHNEIDER andPaul SHRIVASTAVA
87/30 Jonathan HAMILTONV. Bentley MACLEODand J. F. THISSE
87/31 Martine OUINZII andJ. F. THISSE
87/32 Arnoud DE MEYER
87/33 Yves DOZ andAmy SRUEN
87/34 Kasra FERDOVS andArnoud DE MEYER
87/35 P. J. LEDERER andJ. F. THISSE
87/36 Manfred KETS DE VRIES
87/37 Landis GABEL
87/38 Susan SCHNEIDER
87/39 Manfred KETS DE VRIES
1987
87/40 Carmen NAMES andPierre REGIBEAU
"The internal and external careers: atheoretical and cross-cultural perspective',Spring 1987.
'The robustness of KDS configurations in theface of incomplete date", March 1987, Revised:July 1987.
"Demand complementarities, household productionand retail assortments", July 1987.
"Is there a capital shortage In Europe?",August 1987.
'Controlling the interest-rate risk of bonds:an introduction to duration analysis andimmunization strategies", September 1967.
"Interpreting strategic behavior: basicassumptions themes in organizations", September1987
"Spatial competition and the Core', August1987.
"On the optimality of central places',September 1987.
"German, French and British manufacturingstrategies less different than one thinks',September 1987.
"ik process framework for analyzing cooperationbetveen firms", September 1987.
"European manufacturers: the dangers ofcomplacency. Insights from the 1987 Europeanmanufacturing futures survey, October 1987.
"Competitive location on netvork, underdiscriminatory pricing", September 1987.
"Prisoners of leadership", Revised versionOctober 1987.
"Privatization: its motives and likelyconsequences", October 1987.
'Strategy formulation: the impact of national
culture', October 1987.
"The dark side of CEO succession", November
'Product compatibility and the scope of entry',November 1987
87/41 Cavriel HAVAVINI andClaude VIALLET
87/42 Damien NEVEN andJacques-P. THISSE
87/43 Jean GABSZEVICZ andJacques-F. THISSE
87/44 Jonathan HAMILTON,Jacques-F. THISSEand Anita VESKAMP
87/45 Karel COOL,David JEMISON andIngemar DIERICKY
87/46 Ingemar DIERICKAand Karel COOL
1988
88/01 Michael LAVRENCE andSpyros MAKRIDAKIS
88/02 Spyros MAKRIDAKIS
88/03 James TEBOUL
88/04 Susan SCHNEIDER
88/05 Charles WYPLOSZ
88/06 Reinhard ANGELMAR
88/07 Ingemar DIERICKX
and Karel COOL
88/08 Reinhard ANGELMAR
and Susan SCHNEIDER
88/09 Bernard SINCLAIR-DESCAGN4
88/10 Bernard SINCLAIR-
DESCAGN6
88/11 Bernard SINCLAIR-DESCACNi
"Seasonality, size premium and the relationshipbetween the risk and the return of Frenchcommon stocks", November 1987
"Combining horizontal and verticaldifferentiation: the principle of max-mindifferentiation", December 1987
'Location", December 1987
"Spatial discrimination: Bertrand vs. Cournotin a model of location choice", December 1987
"Business strategy, market structure and risk-return relationships, a causal interpretation",December 1987.
"Asset stock accumulation and sustainabilltyof competitive advantage", December 1987.
"Factors affecting judgemental forecasts andconfidence intervals', January 1988.
"Predicting recessions and other turningpoints", January 1988.
"De-industrialize service for quality', January1988.
'National vs. corporate culture: implicationsfor human resource management", January 1988.
'The slanging dollar: is Europe out of step?",January 1988.
"Les conflits dans les canaux de distribution',January 1988.
"Competitive advantage: a resource based
perspective", January 1988.
"Issues in the study of organizational
cognition", February 1988.
'Price formation and product design throughbidding', February 1988.
'The robustness of some standard auction game
forms", February 1988.
"When stationary strategies are equilibrium
bidding strategy: The single-crossingproperty', February 1988.
88/12 Spyros MAKRIDAKIS
88/13 Manfred KETS DE VRIES
08/14 Alain NOEL
'Business firms and managers in the 21st
century', February 1988
"Alexithysie in organizational life: theorganization man revisited', February 1988.
me interpretation of strategies: a study ofthe impact of CEOs on the corporation',March 1988.
88/29 Naresh K. MALHOTRA,Christian PINSON andArun K. JAIN
88/31 Sumantra CMOS/1AL andChristopher BARTLETT
'Consumer cognitive complexity end thedimensionality of aultidisensional scalingconfigurations', May 1988.
'Creation, adoption, and diffusion ofinnovations by subsidiaries of multinationalcorporations', June 1988.
88/30 Catherine C. ECKEL "The financial fallout from Chernobyl: riskand Theo VERMAELEN perceptions and regulatory response', May 1988.
88/24 B. Espen DC1010 and
Hervig LANGOHR
'The production of and returns froe industrialinnovation: an econometric analysis for a
developing country', December 1987.
'Market efficiency and equity pricing:international evidence and laplications forglobal investing', March 1988.
"Monopolistic competition, costs of adjustmentand the behavior of European employment',
Septenber 1987.
'Reflections on *Veit Unemployment' inEurope", November 1987, revised February 1988.
'Individual bias in judgements of confidence',March 1988.
'Portfolio selection by mutual funds, anequilibrium model", March 1988.
'De-industrialize service for quality',March 1988 (88/03 Revised).
'Proper Quadratic Functions with an Applicationto MDT", May 1987 (Revised March 1988).
'Equilibres de Nash-Cournot dans le arch!
europien 4u gel: un cis 6 les solutions enboucle ouverte et en feedback coincident',
Mars 1988
'Information disclosure, means of payment, and
takeover premix. Public and Private tenderoffers In Prance', July 1985, Sixth revision,
April 1988.
88/15 Anil DEOLALIKAR endLars-Hendrik ROLLER
88/16 Gabriel HAVAVINI
88/17 Michael BURDA
88/18 Michael BURDA
88/19 M.J. LAVRENCE andSpyros MAKRIDAKIS
88/20 Jean °ERMINE,
Damien NEVEN andJ.F. TIIISSE
88/21 James TEBOUL
88/22 Lars-Hendrik ROLLER
88/23 Sjur Didrik FLAN
and Georges 2.ACCOUR
88/32 Kasra FERDOVS and
'International manufacturing: positioningDavid SACKRIDER
plants for success', June 1988.
88/37 Mihkel M. TOMBAK
'The importance of flexibility inmanufacturing', June 1988.
88/34 Mihkel M. TOMBAK
'Flexibility: en important dimension inmanufacturing', June 1988.
88/35 Mihkel M. TOMBAK
"A strategic analysis of investment in flexibleeanufacturing systems', July 1988.
88/36 Vlkas TIBREVALA and 'A Predictive Test of the N80 Model thatBruce BUCHANAN
Controls for Non-stationarlty", June 1908.
08/37 Murugappa KRISHNAN 'Regulating Price-Liability Competition ToLars-Hendrik ROLLER Improve Velfere', July 1988.
88/38 Manfred KETS DE VRIES 'The Motivating Role of Envy : A ForgottenPatter in Management, April 88.
88/39 Manfred PETS DE VRIES "the Leader as Mirror : Clinical Reflections",
July 1988.
88/40 Josef LAKONISBOK and 'Anomalous price behavior around repurchaseTheo VERMAELEN tender offers', August 1988.
88/41 Charles VIPLOSZ
• Assysetry in the ENS! intentional or
systesicI", August 1988.
88/42 Paul EVANS
'Organisational development in thetransnational enterprise', June 1988.
88/43 B. SINCLAIR-DESGAGNE 'Croup decision support systems implementRayesian rationality', September 1988.
88/25
Everette S. GARDNER
and Spyros MAKRIDAKIS
88/26 Sjur Didrik FLAP'
and Georges ZACCOUR
88/27 Murugappa KRISHNAN
Lars-Rendrik ROLLER
88/28 Sunantra GHOSRAL andC. A. BARTLETT
'The future of forecasting', April 1988.
'Semi-competitive Cournot equilibrium in
multistage oligopolies', April 1988.
'Entry tame with resalable capacity',
April 1988.
'The multinational corporation as a netvork:
perspectives from interorganizational theory'',
May 1988,
88/44 Essaa MAHMOUD andSpyros MAKRIDAKIS
88/45 Robert KORAJCZYKand Claude VIALLET
88/46 Yves DOZ andAmy MEN
'The state of the art and future directionsin combining forecasts', September 1988.
'An empirical investigation of internationalasset pricing', November 1986, revised August1988.
'From intent to outcome: a process frameworkfor partnerships', August 1988.
88/47 Alain INATEZ,Els CIJS8RECRTS,Philippe NAERT and/let VANDEN A_AEELE
88/48 Michael BURDA
88/49 Nathalie DIERKENS
88/50 Rob VEITZ andArnoud DC MEYER
88/51 Rob VEITZ
88/52 Susan SCHNEIDER endReinhard ANGELMAR
88/53 Manfred KITS DE VRIES
88/54 Lars-Hendrik RULERand Mlhkel M. TOMBAK
88/55 Peter BOSSAERTSand Pierre MILLION
88/56 Pierre MILLION
88/59 Martin KILDUFF
88/60 Michael BURDA
88/61 Lars-Rendrik ROLLER
88/62 Cynthia VAN RUE,Theo YERMAELEN andPaul DE vouttas
'Asywaetslo cannibalise between substitutekite, listed by retailers'. September 1988.
"Reflections on 'Veit unesploysent' inEurope, II', April 1988 revised September 1988.
"Information asymmetry and equity issues*,September 1988.
'Managing expert systess: from inceptionthrough updating", October 1987.
"Technology, work, and the organisation) thelopact of expert systems", July 1988.
'Cognition and orgenisationel mmalyaielminding the store', September 1988.
'Vhatever happened to the philosopher - kings theleader's addict ion to power, September 1988.
'Strategic choice of flexible productiontechnologies and velfare implications',October 1988
'Method of moments tests of contingent claimsasset pricing models', October 1988.
'Sire-sorted portfolios and the violation ofthe random valk hypothesise Additionalempirical evidence and Implication for testsof asset pricing models', June 1988.
'The interpersonal structure of decisionmaking: a social cosparison approach toorganisational choice', November 1988.
'Is mismatch really the problem/ Some estimatesof the Chelvood Cate II model vith US data',September 1988.
"Modelling Cost structures the sell Systemrevisited', November 1988.
'Regulation, taxes and the market for corporatecontrol in Belgium', September 1988.
88/63 Fernando NASCIMENTOand Wilfried R.VANHONACKER
88/64 Kasra FERDOWS
88/65 Arnoud DE MEYERand Kasra FERDOWS
88/66 Nathalie DIERKENS
88/67 Paul S. ADLER andKasra PERDOWS
1989
89/01 Joyce K. BYRER andTavfik JELASSI
89/02 Louis A. LE BLANCand Tavfik JELASSI
89/03 Beth H. JONES andTavfik JELASSI
89/04 Kasra PERDOVS andArnoud DE MEYER
89/05 Martin KILDUFF andReinhard ANCELMAR
89/06 Mihkel M. TOMBAK andB. SINCLAIR-DESGAGNE
89/07 Damien J. NEVEN
89/08 Arnoud DE MEYER andHellmut SCHUTTE
89/09 Damien NEVEN,Carmen MATUTES andMarcel CORSTJENS
89/10 Nathalie DIERKENS,Bruno GERARD andPierre BILLION
"Strategic pricing of differentiated consumerdurables in a dynamic duopoly: a numericalanalysis", October 1988.
"Charting strategic roles for internationalfactories", December 1988.
"Quality up, technology dovn", October 1988.
"A discussion of exact measures of informationassymetry: the example of Myers and Majlufmodel or the importance of the asset structureof the fir.", December 1988.
"The chief technology officer", December 1988.
"The impact of language theories on DSSdialog", January 1989.
"DSS softvare selection: a multiple criteriadecision methodology", January 1989.
"Negotiation support: the effects of computerintervention and conflict level on bargainingoutcome", January 1989."Lasting improvement in manufacturingperformance: In search of a nev theory",January 1989.
"Shared history or shared culture? The effectsof time, culture, and performance oninstitutionalization in simulatedorganizations", January 1989.
"Coordinating manufacturing and businessstrategies: I", February 1989.
"Structural adjustment in European retailbanking. Some viev from industrialorganisation", January 1989.
"Trends in the development of technology andtheir effects on the production structure inthe European Community", January 1989.
"Brand proliferation and entry deterrence",February 1989.
"A market based approach to the valuation ofthe assets in place and the grovthopportunities of the fir.", December 1988.
88/57 vilfried VANHONACKER 'Clete transferability: esti.ating the responseand Lydia PRICE effect of future events based on historical
analogy", October 1988.
88/58 B. SINCLAIR-OCSGAGNE 'Assessing econoalc inequality', November 1988.and Mlhkel M. TOMBAK
89/11 Manfred KETS DE VRIESand Alain NOEL
89/12 Wilfried VANHONACKER
89/13 Manfred KETS DE VRIES
89/14 Reinhard ANGELMAR
89/15 Reinhard ANGELMAR
89/16 Wilfried VANHONACKER,
Donald LEHMANN andFareena SULTAN
89/17 Gilles AMADO,Claude FAUCREUX andAndre LAURENT
89/18 Srinivasan BALAK-RISHNAN andMitchell KOZA
"Understanding the leader-strategy interface:application of the strategic relationshipinterview method", February 1989.
"Estimating dynamic response models when thedata are subject to different temporalaggregation", January 1989.
"The impostor syndrome: a disquietingphenomenon in organizational life", February1989.
"Product innovation: a tool for competitiveadvantage", March 1989.
"Evaluating a firm's product innovationperformance", March 1989.
"Combining related and sparse data in linearregression models", February 1989
"Changement organisationnel et realtiesculturelles: contrastes franco-americalns",March 1989
"Information asymmetry, market failure andjoint-ventures: theory and evidence",March 1989