with moment conditions
E(sit∖zit-ι) = E(μit-ινiE((wit — Wit-ι)∖νi, Zit-1)∖zit-1) = 0,
where zit-1 is a 1 × P vector of instruments.
In our empirical analysis, we primarily adopt the assumption that busi-
ness tax rates are exogenous to foreign headquarters location. Moreover, we
mainly exploit variation contained in the cross-section, since the time series
is short for which information about the municipality-level variables is avail-
able for all cross-sectional units (ranging from 2001 to 2005 at most). Over
a short time-span, it is difficult to discern effects of explanatory variables
with small time series variation from fixed municipality effects. However, we
apply the above instrumental variable techniques with cross-sectional and
panel data to shed light on the robustness - in qualitative terms - of the
cross-sectional business tax effect on foreign MNE headquarters location.
5 Empirical analysis
We start by estimating models on the cross-sectional data-set and assume
that the N × K matrix of explanatory variables includes - apart from TAX
(the log of the business tax rate) - the following variables: SKILL (the log
share of employees with a tertiary education), POPDEN (the log population
density), AREA (the log total area of the municipality in square kilometers),
I DEP RAT (the log independency ratio, defined as the population aged 15-
64 divided by the total population), BUILT (the log fraction of area in
a municipality which is covered by buildings and streets), INV (the log
share of investment expenditures of the municipality in total expenditures),
and EAST (a dummy for municipalities located in the New Lander, i.e., in
the former Eastern German part of the country). Table 2 gives summary
statistics for the covariates included in the analysis.21
In the cross-sectional models, we use two alternative dependent vari-
ables. One measure is the stock of MNEs in a municipality as of 2005 (la-
beled MNE) and the other one the number of new headquarters in 2005
(NEWMNE). The latter may help avoiding a bias of the model parameters
arising from the possible correlation between the tax rate of a municipality in
21The statistics correspond to the pooled observations over the period 2001-2005. The
statistics for the cross section of 2005 are virtually identical.
14