| Civilian | Military | |
|---|---|---|
| Personalistic | Boss | Strongman |
| Non-personalistic | Machine | Junta |
| Note: | ||
| Based on Figure 1 in Weeks (2012) |
14 Checks and Balances
Blattman (2023) argues that checks and balances are an effective means to preserving peace. The most obvious case where this is true is with respect to leader interests. Unconstrained leaders may not internalize the costs of war felt by the broader population, but when political and material power is distributed across multiple branches of government, bureaucratic agencies, and informal institutions a leader is forced to contend with the interests of a wider swath of society before making the decision to go to war. How well does this argument shake out when we look at the data?
Before we go too far down the rabbit hole, we first should probably address the fact that this research question feels like deja vu. In the chapter on Unchecked Interests we talked about how unconstrained leaders are particularly likely to take their countries to war. In particular, we saw that countries with higher democracy quality were better at constraining leader willingness to use militarized force. So question answered, right? Not necessarily. The argument Blattman (2023) makes is not contingent on democracy to create sufficient checks and balances to curtail violence. Rather, what is necessary is polycentric governance—a form of governance where there are many centers of power within society. Non-democracies are capable of exhibiting polycentric governance in addition to democracies (Contemporary China is a good example). Furthermore, democracies can also exhibit highly centralized (monocentric) governance (Liberia before a military junta seized control).
So in this chapter, we’re going to go beyond a simple dichotomy or one-dimensional continuum of autocracy to democracy. In a highly influential study, the political scientist Jessica Weeks (2012) shows that not all autocratic regimes are created equal when it comes to international conflict. Depending on their structure, some autocracies are just as effective at constraining their leaders as democracies. In this chapter, we’ll incorporate Weeks’ data into an analysis of conflict initiation and see whether certain kinds of authoritarian rule really can constrain leaders just as well as in democracies.
14.1 Hypotheses
Weeks (2012) breaks down authoritarian regimes along two dimensions. The first is personalism and the second is militarism. On the one hand, political power can be concentrated in a single leader (high personalism) or a leader can be constrained by an elite “audience” (what others might call a “selectorate”). On the other hand, a country can either be run by a civilian or military regime. Using these dimensions, Weeks creates a four-category typology of authoritarian regimes, as summarized in the below table: (1) boss, (2) machine, (3) strongman, and (4) junta.
Boss and machine regimes have civilian rule, while strongman and junta regimes have military rule. Furthermore, boss and strongman regimes have personalistic forms of rule where power is concentrated in a single leader. Meanwhile, machines and juntas have more distributed power.
Weeks predicted that machines are the most peaceful among these kinds of authoritarian regimes. In some ways, this prediction contradicts some conventional wisdom. Plausibly, an authoritarian leader should have a better chance of convincing a small selectorate of elites of the benefits of going to war than would be the case if this leader had to convince a democratic electorate. However, for this argument to hold up, it must be the case that elites are sufficiently insulated from the costs of war. While this might be true with respect to the direct costs of fighting on the front-lines, there are other costs of war (like economic decline) that these elites may be especially sensitive to. So, as Weeks hypothesizes:
Hypothesis 1: Machines are no more likely to go to war than democracies.
Juntas, according to Weeks (2012), are a different story. In these regimes, the audience/selectorate that a leader is beholden to are military elites. As Weeks states, “[m]ilitary officers’ training leads them to view force as a routine and appropriate policy option, to be wary of diplomacy, and to fear the consequences if they do not act” (Weeks 2012, 334). Therefore, she hypothesizes:
Hypothesis 2: Juntas are more likely to go to war than machines or democracies.
When it comes to the remaining two regime types—bosses and strongmen—Weeks argues that the differences are likely to be less substantial. The reasoning is that the characteristics that drive someone to centralize power in themselves are likely to make them especially belligerent, no matter whether their background is civilian or in the military. Therefore, Weeks proposes:
Hypothesis 3: Bosses and strongmen are more likely to go to war than machines and democracies.
Where there are substantial differences between these two regimes and juntas is not so clear according to Weeks. We’ll ignore fourth hypothesis where she tries to differentiate these regimes from juntas because even her own discussion suggests a considerable amount of uncertainty on her part. The above three give us more than enough to consider.
However, we can go beyond these three hypothesis by considering a variable we discussed several chapters ago for unchecked interests—leader willingness to use force. When Weeks (2012) did here analysis more than 10 years ago, fine-grained data at the leader-level of analysis was not very common or, for that matter, fully baked. However, {peacesciencer} gives us access to this data, which can helps us refine the analysis done by Weeks to consider the interests of leaders in the face of the possible constraints that different regime types impose on them.
Constraints in the face of leader interests are important to consider because, independent of regime type, it is relevant whether a leader is willing to use force if we want to fully understand the probability that a country starts a war. So, in the analysis that follows, we’ll extend the above hypotheses to consider how different kinds of regimes suppress the effect of leader willingness to use force on conflict initiation. Assuming the arguments that Weeks (2012) makes are true, we should expect the following to be true:
Hypothesis 1 becomes \(\to\) Machines are just as capable as democracies of suppressing the effect of leader willingness to use force.
Hypothesis 2 becomes \(\to\) Juntas are not as capable as machines and democracies at suppressing the effect of leader willingness to use force.
Hypothesis 3 becomes \(\to\) Bosses and Strongmen are not as capable as machines and democracies at suppressing the effect of leader willingness to use force.
14.2 Data
To test these hypotheses, we’ll construct a leader-year dataset using {peacesciencer}, which we’ll populate with data on MID initiation and peace spells. We’ll also include details about democracy and leader willingness to use force. We won’t use the democracy data in the main analysis, but it’s good to include it so that we can cross-reference it with the regime data from Weeks (2012). It’s always a good idea to do a sanity check with our data to ensure that its capturing what we think it’s capturing. If Weeks had accidentally coded some countries with a high democracy score as military juntas, that would be problematic.
Speaking of Weeks’ data, we’ll have to access that and incorporate it into our analysis dataset the “old fashioned way”—that is, go to the author’s website and hope they were kind enough to make their data publicly available. Thankfully, Weeks did this! You can download the data from here 2012 study using the following url: https://www.jessicalpweeks.com/uploads/3/8/8/8/38885863/weeksapsr2012.zip
After I downloaded the data, I saved it in my “Data” folder in my working directory in RStudio. You can see the relevant file path on my computer below. Yours may not be the same when you download the data, so make sure you adjust your code accordingly. The relevant data file is called WeeksAPSR2012.dta. This is a Stata data file, which you can read into R using read_dta() from the {haven} R package.
The code below sets up my R environment with some packages I need and creates, by the end, a dataset called Data that has all the relevant variables I need for the analysis.
## packages
library(tidyverse)
library(peacesciencer)
library(coolorrr)
set_theme()
set_palette(
qualitative = c("steelblue", "gray20", "red3")
)
## build country-year data
create_leaderyears(
subset_years = 1945:1999,
standardize = "cow"
) |>
add_gml_mids() |>
add_spells() |>
add_lwuf() |>
add_democracy() -> Data
## read in data from Weeks (2012)
haven::read_dta(
here::here(
"Data",
"WeeksAPSR2012",
"WeeksAPSR2012.dta"
)
) -> weeks_dt
## collapse it to the country-year level
weeks_dt |>
group_by(year, ccode1) |>
summarize(
democracy = replace_na(unique(democracy_1), 0),
junta = unique(juntajlw_1),
strongman = unique(strongmanjlw_1),
machine = unique(machinejlw_1),
boss = unique(bossjlw_1)
) |>
ungroup() |>
rename(
ccode = ccode1
) -> weeks_dt
## merge the data
Data |>
left_join(
weeks_dt
) -> Data
## one more thing...
## make a regime factor
library(socsci)
Data |>
mutate(
regime = frcode(
democracy == 1 ~ "Democracy",
junta == 1 ~ "Junta",
strongman == 1 ~ "Strongman",
machine == 1 ~ "Machine",
boss == 1 ~ "Boss",
TRUE ~ "Other"
)
) -> DataBefore we get to the analysis, I wanted to check a few things about the variables from Weeks’ dataset. First, I wanted to double-check that they actually correspond to the appropriate level of democracy versus autocracy based on a typical and well-validated measure of democratic quality. Remember that the {peacesciencer} will populate our data with three different such measures using the add_democracy() function (Polity, V-Dem, and Extended UDS). In the below code, I look at how the data breaks down with respect to the Extended UDS measure. This is the “compromise” measure of democracy that factors in a bunch of different measures of democracy to construct a unified democracy score (hence UDS). As you can see in the data viz produced below, it looks like Weeks generally did a good job. She created a multiple category variable of regimes which includes democracy, the four authoritarian types discussed in the previous section, and a generic “other” category. These are non-democracies and other regimes that she wasn’t able to reliably classify on the basis of her typology. The kind of data viz produced below is called a ridge plot. It’s a way of showing the density distribution of a variable, but broken down by different categories. Democracies all have above average scores, while the various authoritarian regimes clearly skew toward having very poor UDS values. The “other” category shows the greatest variance, but still tends toward lower UDS values.
## summarizing UDS scores by regime measures
## from Weeks (2012)
library(ggridges)
Data |>
mutate(
xm_qudsest = (xm_qudsest - mean(xm_qudsest, na.rm = T)) /
sd(xm_qudsest, na.rm = T)
) |>
ggplot() +
aes(
x = xm_qudsest,
y = regime
) +
geom_density_ridges() +
geom_vline(
xintercept = 0,
linetype = 2
) +
labs(
x = "Democracy Scores by Regime Category",
y = NULL,
title = "Regime categories map well to\ndemocracy scores",
caption = "(Democracy measured using Extended UDS)"
) +
annotate(
"text",
x = 0,
y = 1.5,
label = "Sample\nAverage",
hjust = 1.1
)
The next bit of code below produces another ridge plot. This one summarizes the distribution of leader willingness to use force by regime type. There is quite a lot of variance in this variable across different kinds of regimes, but this variable obviously clusters on the lower end for some regimes and the higher end in others. Leaders of democracies and “other” regimes show the lowest levels of willingness to use force. Boss and machines show a great deal of variance, but also tend to skew toward the lower end of the spectrum. Perhaps unsurprisingly, the two regimes that tend to have high levels of leader willingness to use force are the two kinds of military regimes—strongmen and juntas.
## does leader willingness to use force differ
## by regime?
Data |>
ggplot() +
aes(
x = theta2_mean,
y = regime
) +
geom_density_ridges() +
labs(
x = "Willingness to Use Force by Regime Category",
y = NULL,
title = "Strongmen regimes and juntas are especially\nlikely to have leaders with a high willingness to use force"
) 
One last thing before we get to the analysis. What about a simple set of descriptive results about how the rate of conflict initiation varies by regime type? The below code generates a dot plot that shows the percentage of the time in the data we observe a country with a particular regime initiation a new militarized interstate dispute with another country in the international system. Interestingly, the results appear to be consistent with the argument that Weeks (2012) makes. Strongmen and boss regimes are the most likely to start MIDs. Democracies and “other” regimes are the least likely. Juntas and machines fall somewhere in the middle. An important question to ask in the analysis that follows is whether this pattern will hold up with a more rigorous statistical model of conflict initiation.
Data |>
group_by(regime) |>
mean_ci(gmlmidongoing_init) |>
ggplot() +
aes(
x = mean,
y = regime,
xmin = 0,
xmax = mean
) +
geom_pointrange() +
labs(
x = "MID Initiation Rate by Regime Category",
y = NULL,
title = "Boss and Strongman Regimes Initiate More\nMIDs than Other Regimes"
) +
scale_x_continuous(
labels = scales::percent,
n.breaks = 10
)
14.3 Analysis
The finalized versions of the three hypotheses proposed in a previous section require an interactive regression model. The code below opens the packages we need to make appropriate inferences for our regression model and to visualize the results. The statistical model that is estimated is a logit model (like always), which is saved as an object called model_fit. The model includes a cubic peace spells trend and interacts the regime type factor with leader willingness to use force (theta2_mean). This allows us to estimate how the effect of leader willingness to use force on conflict initiation varies by regime type.
library(sandwich)
library(lmtest)
library(sjPlot)
glm(
gmlmidonset_init ~ regime * theta2_mean +
poly(gmlmidinitspell, 3),
data = Data,
family = binomial
) -> model_fitThe results are summarized in the figure produced below. The predicted probability of MID initiation is shown by regime type and also by the level of leader willingness to use force. Three representative values of leader willingness to use force are shown (1 standard deviation below the mean, the mean, and 1 standard deviation above the mean). The results reveal some obvious heterogeneity across regimes. Contrary to the findings of Weeks (2012), it turns out that Machine and Boss regimes are the least capable of constraining leaders that are willing to use force. Conversely, the regimes with a military audience or personalistic leader seem to have better success at constraining leaders. In democracies, juntas, strongman regimes, and other kinds of regimes, lead willingness to use force makes almost no or very little difference in the predicted probability that a country initiates a dispute with another country. But in machines and boss regimes, if the leader is increasingly willing to use force, the probability that they take their country to war increases significantly. What is interesting here is that both of these regimes are civilian rule. We can only speculate about why this would be the pattern we see in the data. I discuss some possibilities in the next section.
plot_model(
model_fit,
type = "int",
terms = c("regime", "theta2_mean"),
vcov.fun = vcovCL(
model_fit,
cluster = "leader",
type = "HC0"
),
mdrt.values = "meansd"
) +
labs(
x = NULL,
y = NULL,
title = "Predicted likelihood of MID initation by\nleader willingness to use force and type of regime",
color = "LWUF"
) +
ggpal(
labels = c("-1 S.D.", "Mean", "+1 S.D.")
)
14.4 Conclusion
The findings above show that democracy is not the only kind of regime able to constrain leader interests. Clearly, some authoritarian regimes are equipped to check their leaders as well. However, the kinds of authoritarian regimes that are best at doing this are not the ones we’d expect. Weeks (2012) argued that strongmen and bosses would be the least equipped to constrain their leaders, but the data only partially support this view. This seems to be true in the case of bosses, but instead of strongmen being uniquely unconstrained, machines are. It appears that the most important mediating factor in whether a country goes to war isn’t whether the regime is personalist versus nonpersonalist. It is whether the regime has a civilian versus military dictatorship, with the civilian dictatorships being especially ill-equipped to constrain their leaders. Why would this be?
First of all, these results might simply be the result of limitations with the data and the statistical model. Weeks (2012) in her own analysis uses a directed dyad dataset to estimate the effect of regime type on MID initiation—a choice that she justifies on the fact that there are many possible factors associated with the strategic environment that countries inhabit that might lead them to be more or less belligerent. Some of these factors could also be correlated with different regime types. Maybe a directed dyadic analysis that controls for different factors would tell a different story.
However, if we take the results at face value, it raises some interesting possibilities. The first is that there could be something unique to civilian dictatorships that makes them less able to constrain their leaders than in military regimes. Maybe these dictorships have an advantage in running propaganda campaigns and using other tactics for whipping a civilian population into a hawkish foreign policy orientation. A military regime, by contrast, may be less secure domestically, which makes it harder to justify a foreign conflict at the risk of losing control of the government at home.
Statistical artifact or not, the idea that civilian dictatorships might actually be more conflict prone than military dictatorships has been proposed in the literature elsewhere(Kim 2018). The problem, of course, is that we’re ultimately dealing with observational data, so strong causal claims are still a far-off reality.