. 10
( 35 .)


I include a number of control variables to account for competing
hypotheses. Most important is a control for social heterogeneity. We
know that social heterogeneity interacts with electoral rules to shape
coordination within electoral districts (Cox 1997), and it is reasonable
to expect that the same will be true across districts. The independent
effect of social heterogeneity on aggregation should be straightfor-
wardly negative “ higher levels of social heterogeneity (especially where
groups are geographically concentrated) should hinder attempts to
aggregate across districts. In terms of in¬‚ation, social heterogeneity
should be associated with higher in¬‚ation scores. However, heteroge-
neity should also modify the effect of the institutional variables on
in¬‚ation. Speci¬cally, high levels of social heterogeneity should reduce
the positive effect of centralization on aggregation. Likewise, for any
given level of social heterogeneity, a greater centralization of authority
should yield better aggregation.
As my proxy for social heterogeneity I use ef “ a measure of ethnic
fractionalization within a country. Clearly ethnic differences are only a
single dimension of overall social heterogeneity, and ethnic fraction-
alization is not a perfect measure of ethnic heterogeneity. Ideally we
might also want information about the polarization of ethnic groups,
their geographic concentration or dispersion, and the extent to which
cleavages are crosscutting or reinforcing.14 Despite these weaknesses,
ethnic fractionalization is a common proxy for social heterogeneity in
the existing literature, and for this reason I choose to use it here. My
measure of ethnic fractionalization comes from Fearon (2003) and is
calculated as 1 À Rgi2 where gi is the percentage of the population of the
ith ethnic group. The ef data are available for 160 countries and

See Selway (n.d.) for a review of different ways of thinking about social heterogeneity.
Testing the Theory 55

represent an improvement on the more common but frequently criti-
cized index of ethno-linguistic fractionalization (EFL) (Posner 2004).
In addition to social heterogeneity, I also control for the age of the
largest government party. One alternative explanation for the degree of
aggregation is that it is simply a function of the age or institutionali-
zation of the party system. In early elections, newly formed parties have
not yet established the reputation or organizational capacity necessary
to effectively run a national campaign. As time passes, some parties
begin to build a reputation and capacity, whereas other, weaker con-
tenders drop out of electoral politics or merge with other parties. Thus
we might expect aggregation to improve along with the age and insti-
tutionalization of the party system. The older and more institutional-
ized the political parties, the better the aggregation.15 Log_govage is
designed to account for this alternative explanation. Log_govage is the
logged age of the largest party in the government coalition. The vari-
able is logged since I expect the marginal effect of a unit increase in a
party™s age on aggregation to decrease as the party grows older. In other
words, a move from 5 to 10 years might be expected to have a bigger
marginal effect than an increase from 50 to 55 years. The information
for this variable comes from the DPI (Beck et al. 2001). (As a robustness
check, I also control for whether a country was an advanced, industrial
democracy, or a developing democracy, using a dummy variable for
whether or not a country was a member of the OECD in 1990. The
substantive ¬ndings are robust to the inclusion of this control.)
As a ¬nal control variable, I include information on the number of
electoral districts across which candidates and parties must coordinate.
The intuition behind this variable is that coordination/aggregation is
more dif¬cult as the number of districts increases. Log_districts is the
log of the number of electoral districts in the lowest electoral tier for the
lower house of the legislature. The variable is logged to re¬‚ect the
expectation of a decreasing marginal effect of a unit increase in district
as the number of districts increases.
As discussed earlier, the dataset I use to test these hypotheses includes
information on 280 democratic legislative elections in 46 countries
covering the period 1970 to 2002. These are countries for which both

On party system institutionalization, see Mainwaring and Scully (1995) and Main-
waring (1999).
Building Party Systems in Developing Democracies

table 3.2. Summary Statistics

Summary Statistics (Cross-Sectional Dataset)
Variable Obs Mean Std. Dev. Min Max
in¬‚ation .16 .14 .55
46 0
senate .60 .49
46 0 1
subrevgdp .21
36 6.28 6.11 21.22
military5 .15 .29
46 0 1
parindex .87 .61
44 0 2
ef .35 .25 .85
44 0
horizontal .85 .63
46 0 2
decentralization .77
36 1.47 0 3
log_govage .88
44 3.68 1.47 5.13
log_districts 46 3.99 1.23 1.79 6.44
probability .15 .20 .75
21 0
ENPres 36 1.98 1.77 0 5.5
proximity .28 .40
44 0 1
log_avemag 44 1.31 1.16 0 5.01
ENPnat 46 3.92 1.81 1.95 9.40
Summary Statistics (Pooled Dataset)
Variable Obs Mean Std. Dev. Min Max
in¬‚ation .15 .15 .70
senate .69 .46
273 0 1
subrevgdp .13
205 8.25 6.42 23.44
military5 .12 .32
271 0 1
parindex .83 .61
208 0 2
ef .30 .23 .85
263 0
horizontal .81 .51
271 0 2
decentralization .62
203 1.36 0 2
log_govage .69
213 3.73 1.04 5.16
log_districts 271 4.0 1.27 1.79 6.46
probability .16 .23
141 0 1
ENPres 195 1.67 1.7 0 6.57
proximity .20 .40
265 0 1
log_avemag 264 1.45 1.21 0 5.01
ENPnat 273 3.97 1.91 1.63 13.79

district level and national level election data were available.16 Each
country case contains nearly six elections on average, ranging from a low

A project, funded by the National Science Foundation, to extend the district return
dataset to more elections in more countries is currently underway.
Testing the Theory 57

of one election in ¬ve countries to 15 in the United States. See Table 3.1
for a complete list of the countries and their in¬‚ation scores. Table 3.2
contains the summary statistics for the dataset.
I test my hypotheses using both cross-sectional and pooled analyses. In
the pooled analyses, I use OLS with robust standard errors clustered by
country. This is a better modeling option than employing a ¬xed effects
model with panel-corrected standard errors (PCSE) (Beck and Katz
1995) given the nature of the data (Franzese 2006; Golder 2006). First,
¬xed effects minimize the explanatory power of my time invariant in-
dependent variables. Clustering the standard errors by country allows me
to produce consistent estimates of the standard errors while accounting
for unit-heterogeneity in a way that does not require ¬xed effects
(Franzese 2006; Golder 2006). Second, the accuracy of PCSEs
increases as the number of observations per unit increase. Because
many countries have only a few elections represented in the dataset,
we might question the advisability of using PCSEs. Note, however,
that while I report only results using clustered standard errors, using
PCSEs instead does not change the interpretation of the results.17
To correct for serial correlation in longitudinal data, it is also
common to include a lag of the dependent variable on the right-hand
side. However, as Golder argues in a recent study on elections in presi-
dential democracies, using a lagged dependent variable with compara-
tive electoral data is problematic (2006). “First, observations in the
dataset do not always come in regular intervals either within countries or
across countries.” (Golder 2006, 9) For example, the period between
Thai elections ranges from 1 year to as many as 5 years. Given this
irregularity, it is dif¬cult to know how one would interpret the estimated
coef¬cient on a lagged dependent variable. Second, the panel nature of
the dataset (small T, large N) means that including a lagged dependent
variable would signi¬cantly reduce the sample size and drop all countries
for which I have data on only a single election (Golder 2006, 9). For these
reasons, the models I present here do not include a lagged dependent

In fact, in almost every case, the use of PCSEs yields stronger results.
Including the lagged dependent variable does not generally change the nature of my
inferences. With the lagged dependent variable included, the signs of other explan-
atory variables generally remain the same, though in some cases the variables are no
longer statistically signi¬cant.
Building Party Systems in Developing Democracies

table 3.3. Horizontal Centralization (Dependent Variable: In¬‚ation)

Cross-Sectional Analyses Pooled Analyses
5 6
Variable 1 2 3 4
senate 0.06 0.03 0.08** 0.08**
(0.04) (0.05) (0.03) (0.04)
military5 0.24** 0.28** 0.19*** 0.16**
(0.07) (0.10) (0.05) (0.07)
parindex 0.07** 0.02 0.06* 0.05
(0.03) (0.03) (0.03) (0.03)
subrevgdp 0.004
0.007* 0.007* 0.003
(0.004) (0.004) (0.003) (0.003)
horizontal 0.10***
(0.04) (0.03)
ef 0.33***
0.33*** 0.34*** 0.31***
(0.10) (0.11) (0.11) (0.11)
À0.04 À0.08*** À0.04**
(0.03) (0.03) (0.02)
log_districts 0.002
0.005 0.01
(0.018) (0.02) (0.03) (0.02)
Constant 0.10
0.03 0.09 0.21* 0.03 0.11
(0.04) (0.13) (0.12) (0.04) (0.10) (0.09)
R-squared 0.46
0.29 0.59 0.49 0.26 0.49
Observations 161
44 34 34 208 146
* p < .10, **p < .05, *** p < .01; standard errors in parentheses

3.3.1 Results

Table 3.3 displays the results from six different simple additive models.
Columns 1“3 display the results using the cross-sectional analyses.
Columns 4“7 display the results from the same models run on the
pooled dataset. The dependent variable in all of the models is the
in¬‚ation score. Starting with models 1 and 2 and the corresponding
models 4 and 5, the results in Table 3.3 provide strong support for the
reserve domains hypotheses and some support for the bicameralism
hypothesis as well. The presence of reserve domains, as measured by
military5, signi¬cantly increases party system in¬‚ation in all four
models. In two of the four models, bicameralism is signi¬cantly asso-
ciated with poorer aggregation (as manifest by an increase in the
in¬‚ation score). In the two remaining models, bicameralism has the
Testing the Theory 59

correct sign but is not statistically signi¬cant. There is also strong
statistical support for two of the three control variables. As expected,
greater ethnic fragmentation is associated with poorer aggregation
(higher in¬‚ation), whereas party systems with older parties appear to
do a better job aggregating across districts than those with younger
parties. The number of districts, however, appears to have no signi¬-
cant affect on the degree of cross-district coordination.
Fairing less well is my measure of party factionalism “ parindex.
Parindex is signi¬cant in two of the four models (and it does have the
correct sign in the rest). However, even though my proxies for reserve
domains and bicameralism are robust to alternative speci¬cations,
parindex is not. Across all the various models parindex is rarely sig-
ni¬cantly related to in¬‚ation, and in some models switches signs.
Parindex does not perform any better if I recode it as dummy variable.
In short, parindex does not appear to be a good proxy for party fac-
tionalism “ not completely surprising given that it was at best a rather
indirect measure of party cohesion.19 In addition, including parindex
in the models comes at a cost since it is available for fewer years than
the rest of the sample.20 Speci¬cally, including parindex reduces the
sample size by more than 23%. Both because of the variable™s poor
performance and in a desire to economize on observations, I opt to
drop parindex from the remainder of the analyses in this chapter.
However, I will revisit the party factionalism hypothesis in Chapter 5
in connection with the Thai case.


. 10
( 35 .)