includes player fixed effects, we do not include separate age covariates, or interactions
between euron and the age variables. The interactions with af tert pick up any particular
relationship between age and output after the qualification date but not specific to players
from qualified nations.
Because our data on minutes played take on nonnegative integer values (between 0 and
96), a count model is appropriate for regressions with minutes played as the dependent
variable; we will use the negative binomial model, since the Poisson model is rejected at
high degrees of confidence.31 For all other dependent variables, which are outputs (shots
on goal,...) per minute played and hence continuous, we use OLS estimation. While all the
stated regression equations in principle allow error terms to depend on n, within-group
correlations at the nationality level are low.32 This is not surprising, since as already
argued it is difficult to come up with any reason - unrelated to the national team - for
why nationality should matter in an established player’s professional life. Standard errors
are robust and clustered at the individual player level so as to take into account serial
correlation, which could be due for example to injuries.33
5 Results
5.1 Average Effects
Before turning to the regression results, we present graphical evidence of trends in the
raw data. Figure 6 tracks the average numbers of minutes played per match of players in
the treatment group, i.e., of Euro 2008 nationalities, and the control group comprising
all other nationalities. For each season and each group the Figure includes a fitted linear
trend, with a 95% confidence interval around it. The two vertical lines indicate the period
during which nations, other than the host countries Austria and Switzerland, de facto
qualified for the Euro 2008. The data exhibit a clear positive trend for players of Euro
31 Allison and Waterman (2002) as well as Guimaraes (2008) show that for the negative binomial model
the estimator proposed by Hausman et al. (1984) is a conditional fixed effects estimator under very
specific assumptions only. As suggested by Allison and Waterman (2002), player fixed effects can be
included by means player dummies, however, which is the approach we follow.
32 For all the output measures we use, the intraclass correlation if class is nationality lies below 0.1, in
many cases even below 0.05.
33If class is player identity, the intraclass correlations for the various output measures we empoy lie
between 0.2 and 0.4.
27