¬gure 3.2. Probability and Decentralization Interacted

do not change much, however, when I exclude Thailand (see Figures 3.2c

and 3.2d). The marginal effect of decentralization is still positive and

increasing in probability, though with Thailand excluded, this marginal

effect is no longer signi¬cant at high levels of probability.

Finally, the ¬nal two graphs in Figures 3.2e and 3.2f show the

marginal effect of probability as decentralization increases. Note that

where there is a high aggregation payoff (decentralization ¼ 0), the

marginal effect of probability on in¬‚ation is indistinguishable from zero.

In other words, a high aggregation payoff can be enough to induce

aggregation even if the probability of capturing that prize is low.

Building Party Systems in Developing Democracies

70

However, as the aggregation payoff declines, an increase in probability

is associated with a rise in party system in¬‚ation. This positive marginal

effect continues to increase as decentralization increases.

To summarize, the results of these models are consistent with both

the Prime Minister and Expected Utility hypotheses. In the additive

models, an increase in the probability that someone other than the

leader of the largest party will capture the premiership is associated

with greater in¬‚ation. (In those same models, the proxy for the

aggregation payoff “ decentralization “ is signi¬cantly positive, as

expected). The interactive models are consistent with my argument

that aggregation incentives are a re¬‚ection of both the size of

the aggregation payoff and the probability of capturing that

payoff. Decentralization, my measure of the aggregation payoff,

has its strongest marginal effect on in¬‚ation when combined with a

high likelihood of not capturing the premiership. In other words,

aggregation is the worst when there is both a low aggregation payoff

and a low probability of capturing that prize, as hypothesized.

3.5 the probability of capturing the

prize “ presidential systems

In presidential democracies, the probability that the largest legislative

party will also capture the presidency is a function of whether presi-

dential and legislative elections are concurrent and the effective number

of presidential candidates. The arguments about the probability of

capturing the prize of government in presidential systems can be

summarized with the following hypotheses.

Concurrency Hypothesis: In¬‚ation is lower when presidential and

legislative elections are concurrent.

ENPresxConcurrency Hypothesis 1: The effect of concurrent elections

on in¬‚ation is conditional on the effective number of presidential

candidates. As the effective number of candidates increases,

the negative relationship between concurrency and in¬‚ation

weakens.

ENPresxConcurrency Hypothesis 2: The effective number of presi-

dential candidates is positively related to in¬‚ation in concurrent

elections.

Testing the Theory 71

In Chapter 2, I also argued that a ban on reelection for the president

should increase the number of presidential candidates, ceteris paribus,

which suggests the following hypothesis.

Reelection Hypothesis: The effective number of presidential

candidates is positively correlated with restrictions on presiden-

tial reelection.

Data on the effective number of presidential candidates (ENPres) and

the concurrency of elections come from Golder (2005).28 To capture the

proximity of legislative and presidential elections, I use a dummy vari-

able that equals 1 if presidential and legislative elections are held in the

same year, 0 if they are not.29 For restrictions on presidential reelection, I

use a variable coded as 1 if the sitting president is not eligible for

reelection (no_reelection). This operationalization is preferable to a

simple dummy variable for whether or not a country places term limits

on a president. Coding the variable for term limits lumps together

countries where presidents face a single term limit (e.g., Mexico or the

Philippines) with countries that allow presidents to serve multiple but

limited terms (e.g., the United States). Since what I ultimately care about

is whether in any given presidential election there is an incumbent

running, no_reelection seems like the logical choice.30 No_reelection is

calculated from data in the DPI (Beck et al. 2001).

I test these hypotheses using both cross-sectional and pooled models.

I again use OLS with robust standard errors clustered by country for the

pooled analyses for the reasons described previously. To test

the hypotheses, I use three different datasets. The ¬rst is a subset of the

dataset described in the previous sections and consists of 81 elections in

18 presidential democracies. I use this dataset to directly test the

hypotheses relating aggregation to concurrency, the effective number

of presidential candidates, and the interaction of these two variables.

Given that this ¬rst dataset is small I also indirectly test the same

three hypotheses using Golder™s legislative elections dataset (2005).

28

The effective number of presidential candidates is calculated by dividing 1 by the sum

of each candidate™s squared vote share: 1/Rvi2.

29

The results are robust to the substitution of a continuous variable measuring the

distance between presidential and midterm elections on a 0 to 1 scale.

30

Even better would be data on whether the incumbent actually runs. Unfortunately

such data have not been gathered in a single dataset to my knowledge.

Building Party Systems in Developing Democracies

72

This dataset covers all democratic legislative elections in the world

from 1946 to 2000 for a total of 784 elections.31 The dependent

variable from this dataset is the effective number of electoral parties at

the national level (ENPnat). Ceteris paribus the effective number of

legislative parties is positively correlated with the in¬‚ation score.32

However, unlike in¬‚ation, ENPnat does not distinguish between the

degree of aggregation across districts and the amount of coordination

within those districts. To control for the district-level effects, I include

the log of average district magnitude (log_avemag) and ethnic frac-

tionalization (ef). We know that these two variables interact to

determine the effective number of parties at the district level (Cox

1997). The district magnitude data come from Golder (2005) and the

ethnic fractionalization data are from Fearon (2003) as described

previously.

To test the reelection hypothesis I use a third data set “ Golder™s

presidential elections dataset. This dataset includes 294 democratic

presidential elections from 1946 to 2000. I focus only on direct election

in presidential democracies and so exclude from my models elections in

hybrid presidential-parliamentary regimes as well as any cases of

indirect presidential elections (e.g., the United States).33 This brings the

number of observations in the dataset to 170. The dependent variable

from this dataset is the effective number of presidential candidates

(ENPres). The main explanatory variable is the presence (or absence)

of a ban on reelection of the sitting president (no_reelection).

As discussed in Chapter 2, the effective number of presidential

candidates (like the effective number of parties) is a product of an

interaction between electoral rules and social structure (Golder 2006).

Ethnic heterogeneity increases the effective number of presidential

candidates only when accompanied by a permissive electoral formula

(namely, majority runoff). I replicate this ¬nding and then add the

31

The total dataset includes 867 elections. This excludes Columbian elections from

1958 to 1970 when there was an agreement between Columbia™s two major parties to

alternate control of government and the share of legislative seats regardless of elec-

toral results. Another 76 elections are dropped from the sample due to a lack of party

vote share data, which are needed to calculate the effective number of electoral

parties.

32

Speci¬cally, holding the average number of parties at the district level constant.

33

Leaving indirect presidential elections in the dataset does not substantively alter the

results.

Testing the Theory 73

no_reelection variable to see whether it has a signi¬cant independent

effect on the effective number of presidential candidates. Ethnic

heterogeneity is measured, as it has been previously, as the ethnic

fractionalization (ef). Runoff is a dummy variable coded 1 if the

presidential election formula is a runoff, 0 otherwise.

3.5.1 Results

Table 3.6 displays the results using my in¬‚ation dataset and Golder™s

legislative elections dataset. In columns 1, 4, and 5, the dependent vari-

able is in¬‚ation. The effective number of electoral parties (ENPnat) is the

dependent variable in columns 2, 3, 6, and 7. Looking ¬rst at the results

from the in¬‚ation models, we can see that, as expected, party system

in¬‚ation is lower where presidential and legislative elections are con-

current “ proximity has a signi¬cant negative effect when ENPres is 0.

Also in line with the ¬rst interaction hypothesis, the de¬‚ationary effect of

concurrent elections diminishes as the effective number of presidential

candidates rises. The coef¬cient on the interaction term proximity*

ENPres is positive and signi¬cant in all of the model speci¬cations. This

relationship holds even when controlling for the size of the aggregation

payoff as in model 5. (Note that decentralization is still positive and

signi¬cant in model 5, even when controlling for the number of presi-

dential candidates and proximity).

The story is similar if we substitute the effective number of electoral

parties (ENPnat) for in¬‚ation as the dependent variable in models 2, 3, 6,

and 7. Models 2 and 6 use my dataset. For a robustness check, I also run

the same speci¬cations using Golder™s larger legislative elections dataset

(models 3 and 7). The substantive results are the same regardless of

which dataset I use. With ENPnat as the dependent variable, my analyses

are similar to recent studies on the effect of presidential election on

legislative fragmentation (Cox 1997; Mozaffar et al. 2003;

Golder 2006). The ¬ndings are consistent with these existing studies.34

Proximity reduces the effective number of electoral parties (proximity is

negative and signi¬cant in all four speci¬cations), but this effect is

conditional on the effective number of presidential candidates. An

increase in the number of candidates undermines the marginal negative

34

See especially Golder (2006, 11“13).

74

table 3.6. Aggregation in Presidential Democracies

Cross-Sectional Analyses “ Dependent Variable Pooled Analyses “ Dependent Variable

Explanatory 1 2 3 (ENPnat) 7 (ENPnat)

Variables (in¬‚ation) (ENPnat) (Golder) (Golder)

4 (in¬‚ation) 5 (in¬‚ation) 6 (ENPnat)

proximity À0.68*** À6.95*** À5.02*** À0.43*** À0.19** À4.30** À3.43***

(0.04) (1.66) (1.66) (0.10) (0.06) (1.57) (1.26)

1.32** 0.61*

ENPres À0.04** À0.14 À0.04 À0.04* À0.04

(0.01) (0.61) (0.60) (0.03) (0.02) (0.39) (0.36)

0.18*** 1.81** 0.28 0.13*** 0.08*** 1.31** 0.65*

proximity*ENPres

(0.02) (0.72) (0.86) (0.03) (0.02) (0.51) (0.37)

0.11 0.74 1.84 0.19 0.49** 2.43**

ef À0.67

(0.07) (2.73) (1.57) (0.11) (0.16) (1.52) (1.16)

log_govage À0.02* À0.05*** À0.09***

(0.01) (0.01) (0.02)

log_districts À0.01 À0.001 À0.03*

(0.01) (0.01) (0.02)

0.69 0.75 0.81*

log_avemag À0.09

(1.09) (0.60) (0.70) (0.41)

0.73 2.56

ef*avemag À0.46 À0.34

(2.12) (0.89) (1.98) (0.83)

0.13**

decentralization

(0.04)

0.44*** 3.65* 2.04 0.42*** 0.31*** 3.19** 1.39

Constant

(0.06) (1.82) (1.40) (0.10) (0.09) (1.42) (1.31)

0.95 0.80

R-squared .56 .63 .77 .64 .39

15 16 26 61 45 74 182

Observations

* p < .1, ** p < .05, *** p < .01; standard errors in parentheses

Testing the Theory 75

effect of proximity on the number of parties (the interaction term,

proximity*ENPres, is positive in all four speci¬cations, signi¬cantly

so in three).

In Figure 3.3, I display the marginal effect graphs for four of the

seven models in Table 3.6, speci¬cally models 1, 2, 4, and 6.35 The

graphs tell a remarkably consistent story. Concurrent presidential

elections are associated with better aggregation (fewer parties and less

in¬‚ation), but only where there is a small effective number of presi-

dential candidates. Once the number of candidates rises to somewhere

between 2.5 and 3.5, the marginal effect of concurrency disappears

altogether. Focusing for a moment on just in¬‚ation (Figures 3.3a and

3.3c), we see that once the effective number of candidates is suf¬ciently

large, concurrent elections are actually associated with a signi¬cant

increase in in¬‚ation.

Table 3.7 presents the remainder of the presidential election models.

In the ¬rst cross-sectional speci¬cation and models 4 and 5 in the pooled

speci¬cations, I isolate the impact of the effective number of presidential

candidates on in¬‚ation and the number of parties when elections are

concurrent (proximity ¼ 1).36 We can see that when presidential and

legislative elections are concurrent, the effective number of presidential

candidates is positively related to both party system in¬‚ation and the

effective number of electoral parties, as hypothesized. If the effective

number of candidates is in fact so important for shaping aggregation

incentives, as it appears to be, then what determines the effective number

of candidates?

In models 2 and 6, the effective number of candidates (ENPres) is

the dependent variable, and I test whether reelection restrictions lead to

more presidential candidates, controlling for the strength of the elec-

toral system and social heterogeneity. The results provide some support

for the hypothesis that restrictions on reelection are associated with

more presidential candidates. In model 2, no_reelection is positive but

just short of signi¬cant, while in the pooled analysis (model 6) the

coef¬cient for no_reelection is both positive and signi¬cant. The lack of

an incumbent does appear to have an in¬‚ationary effect on the effective

35

The graphs for the other three models tell the same story.

36