. 11
( 35 .)


Recall that bicameralism and reserve domains are hypothesized to
work together to affect the degree of horizontal centralization. In an
attempt to capture this synthesis, I create an index labeled horizontal,
which additively combines senate and military5 to create a variable
that ranges from 0 to 2. Lower scores on horizontal are associated
with a greater concentration of power. In other words, we can think of
horizontal as representing the degree of horizontal decentralization.
In models 3 and 6 in Table 3.3, we see that horizontal is positively
related to in¬‚ation as expected “ greater horizontal decentrali-
zation is associated with higher party system in¬‚ation (poorer

John Carey has developed a possible measure of party cohesion based on legislative
voting records (Carey 2007). Unfortunately the dataset it is not available for enough
of my sample to allow me to use it here.
Parindex begins in 1978 while the rest of the sample begins in 1970.
Building Party Systems in Developing Democracies

aggregation).21 Speci¬cally, the results from Table 3.3 model 6
suggest that a one-unit increase in horizontal (from 0 to 1 or from 1 to
2) is associated with an increase in the predicted in¬‚ation score of .10
(recall that the in¬‚ation score ranges from 0 to 1).22 When the power
is highly centralized horizontally (horizontal ¼ 0) the predicted in-
¬‚ation score is .09. As power becomes more decentralized, in¬‚ation
rises to .19 (when horizontal is 1) and then to .29 (horizontal is 2). By
contrast, a one standard deviation increase of subrevgdp (my proxy
for vertical decentralization) causes only a .02 point increase in the
in¬‚ation score (from .17 to .19). Table 3.9 displays these predicted
in¬‚ation scores as well as predicted values for select other models
from this chapter. Given that horizontal includes only two variables,
it may appear more sensible to continue to include these variables in
the model separately or to interact them. However, it is useful to have
a single summary measure of horizontal centralization, especially
when I turn to a set of interactions in subsequent models.

The Size of the Prize: Vertical and Horizontal Interaction
Table 3.3 by itself provides only weak support for the vertical cen-
tralization hypotheses “ my measure of vertical (de)centralization is
signi¬cant only half of the time. However, remember that one of the
central arguments in Chapter 2 was that the payoff to being the largest
party is a joint product of both vertical and horizontal centralization. A
suf¬ciently high level of vertical centralization may be necessary to
induce cross-district coordination, but absent an adequate degree of
horizontal centralization, it is not suf¬cient. This suggests the following
conditional hypothesis:

Joint Centralization Hypothesis: The negative effect of vertical
centralization on in¬‚ation is conditional on the degree of hori-
zontal centralization. In¬‚ation should be highest where there is
both a high degree of vertical and horizontal decentralization
and lowest where there a high degree of vertical and horizontal

In the cross-sectional analysis, military5 does the bulk of the work when the model is
fully speci¬ed.
Holding all other variables at their means.
Testing the Theory 61

In Table 3.4 I attempt to test this hypothesis by interacting
subrevgdp with horizontal. Columns 1 and 3 summarize the results for
the cross section and pooled analysis. As hypothesized when there is a
high degree of horizontal centralization (horizontal is 0) an increase in
the percentage of revenues controlled by subnational actors (sub-
revgdp) reduces aggregation (increases in¬‚ation). Speci¬cally, a one
standard deviation in subrevgdp when horizontal is 0 boosts the pre-
dicted in¬‚ation score from .06 to .11 (see Table 3.9).23 Likewise when
there is perfect vertical centralization (subrevgdp equals 0) greater
horizontal decentralization is associated with poorer aggregation. The
substantive effect is quite large “ an increase of horizontal from 0 to 1
causes a .14 increase in the in¬‚ation score (from .01 to .15). As either
variable increases, the marginal effect of the other declines as the
negative coef¬cient on the interaction term shows. In short, the data
demonstrate that even where there is perfect vertical centralization,
horizontal decentralization will still undermine cross-district coordi-
nation. Likewise, when power is concentrated within the national
government, party system in¬‚ation will still occur if subnational actors
control substantial shares of power and resources. Aggregation is at its
highest level when there is both horizontal and vertical centralization.
Note that when using interaction terms, the standard errors on the
interaction term and the constituent variables are uninterpretable
(Brambor, Clark, and Golder 2005; Kam and Franzese 2007). Instead,
the question is the marginal effect of one of the interacted variables over
the range of the other variable, and whether that effect is signi¬cant.
Figure 3.1 graphically illustrates the marginal effect of vertical (de)cen-
tralization on in¬‚ation as the degree of horizontal (de)centralization
changes.24 The solid sloping line shows how the marginal effect changes
as horizontal decentralization increases. We can judge whether this effect
is signi¬cant by drawing 95% con¬dence intervals around the line.
Wherever the con¬dence interval lies completely above or below the
zero line, the marginal effect is signi¬cant. Where it straddles the line
the marginal effect cannot be statistically distinguished from zero.

Again, holding all other variables at their means.
See Kam and Franzese (2007) and Brambor et al. (2005) on the advisability of pre-
senting the results in this manner. I thank Brambor et al. for granting me access to
their STATA .do ¬le, which I use to create all the marginal effect graphs in this
Building Party Systems in Developing Democracies

table 3.4. Horizontal and Vertical Centralization (Dependent
Variable: In¬‚ation)

Cross-Sectional Analyses Pooled Analyses
Variable 1 2 3 4
subrevgdp 0.01** 0.007**
(0.004) (0.003)
horizontal 0.09* 0.14***
(0.05) (0.05)
À0.01 À0.006
horizontal (0.01) (0.004)
decentralization 0.06** 0.07***
(0.03) (0.02)
ef 0.34*** 0.32*** 0.34*** 0.37***
(0.11) (0.10) (0.11) (0.12)
À0.07** À0.06*** À0.04** À0.05***
(0.03) (0.02) (0.02) (0.01)
log_districts 0.01 0.004 0.01 0.003
(0.02) (0.019) (0.02) (0.02)
Constant 0.17 0.21** 0.04 0.13
(0.13) (0.09) (0.09) (0.10)
R-squared 0.50 0.44 0.48 0.45
Observations 34 34 161 161
* p < .1, ** p < .05, *** p < .01; standard errors in parentheses

Figures 3.1a and 3.1b display the marginal effect of subnational
governments™ share of revenues (or the degree of vertical decentrali-
zation) on in¬‚ation. Figure 3.1a illustrates the cross-sectional model,
while Figure 3.1b shows the pooled model. We can see in both ¬gures
that the marginal effect of vertical decentralization on in¬‚ation is posi-
tive and signi¬cant, but only at low levels of horizontal decentralization.
As horizontal power becomes more dispersed, the marginal effect of
vertical centralization is no longer signi¬cant. In short, these two graphs
show that aggregation is highest (and in¬‚ation is lowest) where we have
both vertical and horizontal centralization, as hypothesized. In other
words, vertical and horizontal centralization jointly determine the size of
the payoff for cross-district coordination. Figures 3.1c and 3.1d display
the marginal effect of horizontal decentralization as vertical decentrali-
zation changes. The marginal effect of horizontal decentralization is
positive and signi¬cant and decreases in vertical centralization. In other
95% Confidence Interval

Marginal Effect of Vertical Decentralization on
Marginal Effect of Vertical Decentralization on
(a) (b)
Inflation as Horizontal Decentralization Increases (Pooled)
Inflation as Horizontal Decentralization Increases (XS)





Marginal Effect Vertical

Marginal Effect of Vertical
0 .5 1 1.5 2 0 .5 1 1.5 2
Horizontal Decentralization Horizontal Decentralization
(d) Marginal Effect of Horizontal Decentralization
Marginal Effect of Horizontal Decentralization
on Inflation as Vertical Decentralization Increases (Pooled)
on Inflation as Vertical Decentralization Increases (XS)





Marginal Effect of Horizontal
Marginal Effect of Horizontal

0 .5 1 1.5 2
0 .5 1 1.5 2
Vertical Decentralization
Vertical Decentralization

¬gure 3.1. Marginal Effects of Horizontal and Vertical Decentralization on In¬‚ation
Building Party Systems in Developing Democracies

words, as vertical decentralization increases, the marginal effect of
horizontal decentralization moves toward zero (albeit gradually).
In order to interact the size of the payoff with other variables of
interest as the theory calls for, it is useful to generate a single summary
measure of the size of the payoff. To do this, I convert subrevgdp into a
dichotomous variable, recoding all observations below the median as 0
and the remainder as 1. I then add it to horizontal. The resulting var-
iable, decentralization, ranges from 0 to 3 and represents the combi-
nation of both the vertical and horizontal dimensions of centralization.
Note that higher values on the variable correspond to higher levels of
decentralization. Higher values of decentralization should be associ-
ated with a lower payoff, poorer aggregation, and ultimately, a higher
in¬‚ation score. Columns 2 and 4 in Table 3.4 demonstrate that
decentralization is indeed positively related to in¬‚ation, as expected.25
For the remainder of the analysis I will use decentralization as my proxy
for the size of the aggregation payoff.
To summarize the results from this section, the bicameralism and
reserve domain hypotheses were both supported by the data. Second
chambers and the presence of reserve domains are each associated with
poorer aggregation. The data do not support the party factionalism
hypothesis. This might suggest that the hypothesis is incorrect, or that my
proxy for factionalism, incentives to cultivate a personal vote, does a
poor job of capturing the theoretical concept. The index of horizontal
decentralization (the combination of senate and military5) also produced
the hypothesized effect on aggregation. Higher levels of horizontal
centralization are associated with better the aggregation (and lower
in¬‚ation scores). There was no support in any of the models for the


. 11
( 35 .)