a given year and the number of MNEs located in the municipality in previous
years.22 Of course, NEWMNE has a much larger number of zero entries
than MNE, so using NB or ZI-type models as compared to simple Pois-
son regressions is even more important with NEWMNE. We generally use
once-lagged values of the explanatory variables in the econometric models.
Hence, we employ values for 2004 on the right-hand side of all cross-sectional
models. This is to avoid any bias associated with contemporaneous shocks in
the dependent and the explanatory variables. Since all explanatory variables
except for EAST are in logs, the associated parameters can be interpreted
as elasticities. Table 3 summarizes a set of benchmark regression results.
- Table 3 -
The numbers in Table 3 indicate that the over-dispersion parameter is
significantly different from zero. Hence, the negative binomial model is bet-
ter suited for the data and specification at hand than the Poisson model (see
Winkelmann, 2003; and Cameron and Trivedi, 2006). With the fairly large
fraction of zeros in the dependent variable, a separate modeling of the zero
threshold is recommended. However, the parameter estimates are fairly sta-
ble qualitatively across the estimated models. Especially, when comparing
the estimates of the negative binomial model with its zero-inflated coun-
terpart in Table 3. There, even the 95% confidence intervals around the
coefficients are overlapping for most of the parameters (except for constant,
POPDEN, and AREA).
The key parameter of interest here is the one of TAX. It turns out that
controlling for the suggested determinants of MNE headquarters location
eliminates (most of) the bias obviously present in the unconditional relation-
ship between MNE and TAX portrayed in Figures 3a and 3b. In Table 3, the
estimated parameter is unambiguously negative and statistically significantly
different from zero at the one percent level in all models employing MNE
as the dependent variable, and also with the NB model for NEWMNE.
The parameter of TAX is significantly different from zero at 10% with the
zero-inflated negative binomial (ZINB) regression for NEWMNE.
Note that the effect of TAX on the expected number of headquarters
in a municipality is of considerable magnitude. As mentioned before, the
coefficients should be interpreted as elasticities: a one-percent decline in
22We present and discuss estimates based on instrumental variable regressions and panel
data models below.
15