- Figures 1 and 2 -
Let us first consider the simple correlation between the presence of lo-
cal MNE headquarters and the business tax rate in German municipalities
unconditional on other determinants of location. For this, let us use the cross-
sectional data of 2005 and illustrate the relationship between the number of
MNE headquarters and the business tax rate in simple scatter plots.
- Figures 3a and 3b -
In Figure 3a, we consider the relationship for all municipalities. In Figure
3b, we illustrate it only for municipalities with a positive number of head-
quarters. Obviously, irrespective of which of the two figures we look at, the
unconditional relationship is positive. Do municipalities with higher busi-
ness tax rates attract a larger number of foreign headquarters? This sounds
counter-intuitive. However, conditional on other factors - such as the avail-
ability of skilled workers, region size, a relatively large fraction of population
in working-age, etc. - high business tax rates may well be harmful for head-
quarters location irrespective of the unconditional relationship in Figures 3a
and 3b. We may refer to the source of the positive relationship between
the number of foreign MNE headquarters and the local business tax rate in
Figures 3a and 3b as one of endogeneity of business tax rates - i.e., their cor-
relation with observable or unobservable determinants of the number of MNE
headquarters in a municipality. To shed light on the causal effect of business
tax rates, we now turn to multivariate negative binomial and Poisson model
regressions.
4 Econometric issues
According to the descriptive statistics, the number of foreign MNE headquar-
ters in Germany, our dependent variable, is a count which takes the value
zero in many municipalities. Hence, the distribution function of the depen-
dent variable places probability mass at nonnegative values only. Moreover,
the density function is skewed to the left, and the data are concentrated on
a few small discrete values and intrinsically heteroskedastic with variance
increasing in the mean. This nature of the data likely leads to inconsistent
and certainly to inefficient parameter estimates in linear regression models.
10