the total area, and the share of agricultural area (all in logs) for (i) all mu-
nicipalities within a radius of 0 and 25 kilometers from the center of a given
municipality and for (ii) all municipalities within a radius of 25 and 50 km
from the center of a given municipality. In the panel estimation we use the
averages of the share of area covered with buildings and streets, the share of
agricultural area, the independency ratio, and the skilled labor share for all
neighboring municipalities within a 50 km radius.25
The first column summarizes the findings based on an instrumental vari-
able count data model for the cross-sectional data-set of 2005. Results are
based on the GMM routine described in Section 4.26 The instruments are rel-
evant and pass the over-identification test. The parameters of T AX, SKILL
and POPDEN as well as the variables BUILT and AREA exhibit the same
sign as in Table 3 and are statistically significant. The remaining variables
turn out to be insignificant under IV-GMM. As expected, the parameter
estimate of T AX is negative and higher in absolute value than in the bench-
mark models in Table 3. If some remaining endogeneity were to bias the
coefficients in Table 3 downwards in absolute value (i.e. towards the uncon-
ditional relationship in Figures 3a and 3b), we would expect an instrumental
variables model to raise the point estimate as compared to the models in
Table 3. If the origin of the bias was mainly due to the omission of rel-
evant time-invariant variables, we would expect the bias in Table 3 to be
larger when using the stock of MNE headquarters (MNE) rather than new
headquarters (NEWMNE) as the dependent variable. Indeed, it turns out
25All instruments are measured in logs. Notice that the list of instruments is slightly
different between the cross-sectional and the panel data models. The reason is that we may
only exploit information from time-variant instruments in the panel model. For instance,
AREA and EAST do not vary across the years and can not be used as instruments in
the panel data models. With panel data, we use a municipality’s share of area devoted to
agriculture instead of AREA. This variable exhibits some time variation and is orthogonal
enough to BUILT over time to include it separately (by way of contrast, the two variables
are highly correlated cross-sectionally and should not be included in the models together
in Table 3 to avoid efficiency losses associated with irrelevant instruments). Moreover,
weighted characteristics within a radius of 25 kilometers are strongly correlated over time
with the ones within a radius of 50 kilometers over time. So we exclude the former and
only use the latter in the panel models.
26Notice that the econometric model requires the endogenous variable to be unlimited
in its value range. Obviously, this is not the case for TAX. Using logistically transformed
business tax rates (T AXLOG = ln([tax rate]/[1 - tax rate])) ensures that the endogenous
variable (T AXLOG) is unlimited. However, it turns out that the results are very similar
in a model which employs TAX instead of T AX LOG.
18