Pharmaceutical Detailing Elasticities: A Meta-Analysis - Innovation and Marketing in the Pharmaceutical Industry - page 531

and advertising in the response model, and inclusion of interactions between the
marketing-mix instruments in the response model). For brevity, we will explain the
operationalization of the variables along with their effect on the detailing elasticity
in the next section. The selection of these variables was guided by AMS ( 2010 ) and
earlier meta-analyses in marketing.

18.3.5 Estimation Model and Procedure

We model detailing elasticity as a linear function of the selected independent vari-
ables (determinants) like AMS ( 2010 ). We note there are two levels of variation in
our database, i.e., the 373 personal selling elasticity measurements come from dif-
ferent datasets, and the number of elasticity measurements per dataset varies.
Measurements within a dataset share values on several determinants while they may
differ on other determinants. Since determinants at the measurement level (lower
level) and dataset level (higher level) contribute to variation in personal selling elas-
ticity, but some measurements are not independent within a study, there is a nested
error structure that must be accounted for. Similar to AMS, we use hierarchical
linear model (HLM) estimation (e.g., Bijmolt and Pieters ( 2001 , p. 159) and
Raudenbush and Bryk ( 1992 , p. 440)) in our study. Interested readers are referred to
Bijmolt and Pieters ( 2001 ) for more details about HLM.
We performed several checks to ensure the robustness of our results. First, as
there is no direct diagnostic for multicollinearity in hierarchical linear modeling, we
checked the regression condition index that has a value of 5 which indicates low
multicollinearity. Second, we tested for various plausible interaction effects amongst
all our independent variables and the two focal market characteristics, i.e., stage in
the PLC and geographic setting. Since introducing two interaction effects (one with
stage in the PLC, and one with geographic setting) along with the main effect of
each of the independent variables would introduce severe multicollinearity, we
tested the interaction effects one at a time and retained only the signifi cant ones.
Third, we tested for the effects of several other covariates namely, number of obser-
vations in the dataset , whether or not competitive marketing efforts were explicitly
modeled , interactions between temporal aggregation and the inclusion of lagged
effects , and whether inputs and outputs were measured in monetary (vs. physical )
units . We did not fi nd any of these effects to be signifi cant and hence excluded these
from our fi nal model.

18.4 Estimation Results

The HLM model estimates are reported in Table 18.3 . As highlighted here as well
as in the last column of Table 18.3 , we found 12 signifi cant effects. The overall fi t
of our model to the data ( R 2 = 0.308) is satisfactory considering that we are using our
model for descriptive purposes, and also higher than the model fi ts obtained in