Table 3 provides summary statistics for our variables. Due to a lack of
information on isb, we exclude Belgium from the EU15 countries in our main
data analysis.
[Table 3 about here]
4 Results
We estimate probit specifications by maximum likelihood. The benchmark
model is
Pr (anti = 1|employ, x)=Φ (α + βemploy + x0γ) (8)
where Φ (∙) is the standard normal CDF. We are interested in β's contribu-
tion to the probability.
We present marginal effects in terms of probability evaluated at the mean
of each explanatory variable. For a binary explanatory variable, the figure is
the probability difference between observations with the variable equal to 1
and those with the variable equal to 0. Estimated standard errors are based
on the assumption that observations are not necessarily independent within
each region of a country, but are independent across regions of the country.22
aim is to control for race/ethnicity-based attitudes of respondents and see the direction of
influence, e.g., Agresti (2002).
22 Since the probit-estimated conditional mean function is inconsistent when there is
heteroskedasticity, reporting robust standard errors does not solve the problem of varia-
tion in the variance. One way to tackle this is to model the variance as a function of
covariates, e.g., Wooldridge (2002: 463-465). However, we do not report our results from
such probit models because the results changed dramatically by using different sets of
covariates in modelling the variation in the variance. It was unclear a priori which set of
15