@article{2017:hruschka:functional,
title = {Functional Flexibility, Latent Heterogeneity and Endogeneity in Aggregate Market Response Models},
year = {2017},
note = {Aggregate market response models comprise variables which are defined
as sums (e. g., of sales, of advertising budgets) or averages
(e. g., of prices) across persons or households for a certain store,
retail chain, region, etc. We consider sales or market shares of
brands in one product category as dependent variables and marketing
variables ( e. g. prices, advertising budgets, etc.) as independent
variables. Estimates of the effects of marketing variables which are
either too high or too low can be avoided by functional flexibility,
latent heterogeneity and by taking endogeneity into account. Avoiding
such biases is especially important if we want to derive implications
for marketing decision making. Researchers should switch to flexible
methods as these usually perform much better even for new data than
better known parametric alternatives. Such a switch seems to be quite
feasible as the often heavy data limitations of previous decades have
been overcome, at least in consumer goods marketing. Another
motivation for such a change is offered by the fact that in many
studies functional flexibility turns out to be the most important
factor. We compare three flexible approaches, multilayer perceptron,
splines regression and kernel regression which have been applied to
determine aggregate market response functions. We recommend to use the
multilayer perceptron, which is a form of artificial neural net,
because of better statistical performance, lower computation time and
robustness for extrapolated data. Researchers can avoid problems of
flexible methods such as overfitting and implausible estimation
results (e. g., higher sales at higher prices) by careful selection
of model complexity and by introducing restrictions to ensure, e. g.,
monotone or concave shapes. In models with latent heterogeneity
parameters are not constant, but vary according to a continuous or a
finite mixture distribution. We distinguish two types of latent
heterogeneity, a) heterogeneity within regions or within stores across
consumers, and b) heterogeneity across retail chains or stores. We
observe that in aggregate response modeling continuous mixtures have
been applied more often. Moreover, as a rule continuous mixtures beat
finite mixtures. Most studies with parametric response functions which
also allow for latent heterogeneity obtain a better statistical
performance than related models with constant parameters. Many of
these studies also show that heterogeneous models lead to different
implications for optimal decision making, e. g., optimal pricing.
Therefore researches should regularly investigate latent heterogeneity
if they use parametric aggregate market response models. A marketing
variable is endogenous if it is related to an unobserved factor which
also determines the dependent variable. Endogeneity arises, for
example, if management raises prices or advertising budgets because
unobserved preferences of customers change in favor of the respective
brand. If marketing variables are endogenous, classical methods
provide biased estimates of their effects. Frequently instrumental
variable methods are used to correct these biases, but finding
appropriate instrumental variables is often difficult. Researchers may
bypass these difficulties by turning to instrument-free methods of
which the copula-based approach seems to be most attractive one. If
the assumptions of this approach do not hold, linking marketing
variables to parameters of flexible aggregate response functions
constitutes a viable alternative. If researchers are not interested in
measuring the effects of marketing variables, but instead look on
overall predictive performance, they can safely ignore the endogeneity
problem. On the other hand, functional flexibility and to some degree
latent heterogeneity are important because usually they lead to better
predictions.},
journal = {Marketing ZFP},
pages = {17--31},
author = {Hruschka, Harald},
volume = {39},
number = {3}
}