tional variables used in the corresponding cross-sectional estimates (bicameralism,
investiture vote and constructive vote of no confidence), can no longer be included
since they exhibit too little time variation.24 The results again remind of the re-
sults in Table 2. Larger district magnitude significantly raises the likelihood of
coalition governments, while plurality rule raises the likelihood of single-party
governments. The coefficients are not very precisely estimated. Evidently, there
is considerable time variation in the type of government, which cannot be easily
explained on the basis of the sluggish electoral rule variables.
Columns 4 and 5 report on the structural estimates for the type of government.
In the specification, we thus treat party fragmentation as endogenous, and use
the electoral-rule variables (in first differences) as additional instruments (beyond
one additional lag of the dependent variable and one lag of endogenous party
fragmentation). The estimated coefficients on party fragmentation are significant
with the expected sign: more fragmentation increases the likelihood of coalition
governments and reduces the likelihood of single-party majorities. Moreover, we
cannot reject the over-identifying assumptions on the validity of the instruments,
meaning that the electoral rule variables do not exert a direct effect on the type
of government.
While these panel estimates are not as precise as the cross-sectional estimates,
they still confirm our earlier inference and give further support to the predictions
of the model.
Economic effects of electoral rules Last, we return to the effects of the
electoral rule on government spending with a similar specifications as in the cross
sectional regressions. Table 5 reports on estimates based on the legislature panel
as well as the yearly panel. The dependent variable is always overall government
spending. But in the legislature panel, we always measure the size of government
spending in the last year of the legislature (rather than on average throughout
the legislature), to allow the political variables to exercise their full effect. In the
GMM estimates where the data are differenced, the dependent variable is thus
defined as the change in spending from the end of the previous legislature to the
end of the current one. Since the duration of legislatures varies across countries
and time periods, these regressions include a variable measuring the length of the
legislature, in years. Lagged spending is also included in the specification: one
(differenced) lag in the GMM estimates on the legislature panel, and two (level)
lags in the fixed-effect estimates on the yearly panel.
24 Constructive actually varies over time, but only for Belgium in 1993.
43