9 Intangible Incentives

The next reason for war that Blattman (2023) talks about is intangible incentives. As the name implies, these are incentives that are beyond any material prize sides of a dispute seek to obtain such as land, treasure, or resources. Sometimes groups are motivated to fight because of immaterial incentives like righteous outrage, glory and status, or ideological predilections against certain kinds of comprise.

We can point to a number of real world examples where we can see these incentives at work, but when it comes to measuring the intangible in a systematic way there are some hurdles we need to surmount. How do we measure something that isn’t material? This is a problem that Collier and Hoeffler (2004) tried to solve in their classic article, “Greed and Grievance in Civil War.” As the aptly named article suggests, Collier and Hoeffler (2004) sought to empirically investigate alternative explanations for civil wars.

On the one hand, greed seems like a plausible motivation. Different groups in society want power, wealth, or resources and they may go to violent extremes to get what they want. On the other hand, inequality and egregious human rights violations might spur a violent backlash against a repressive government. Since our focus here is intangible incentives for war, we’ll home in on grievances. While this means we’ll exclude things like glory or ideology from our analysis, focusing on grievances gives us a simple target for what kinds of factors we might want to study in an analysis explaining variation in the break out of civil war.

Let’s get to it.

9.1 Greed vs. Grievance as Causes of Civil War

Measuring an intangible incentive like grievances—injustices or righteous anger—is not straightforward. Short of fielding a survey of would-be combatants or doing interviews, it is hard to measure grievances in a systematic and consistent way across countries. Therefore, most quantitative conflict research doesn’t seek to measure grievances directly. Instead this body of research relies on proxy measures. These are factors, which we can reliably measure, that are plausibly related to the concept we really want to measure. In their classic study, Collier and Hoeffler (2004) try to measure grievance using four different factors:

ethnic fractionalization
Polity index
Gini index
ethnic dominance

The first measure takes a value between 0 and 1 and it quantifies how ethnically diverse a country is. 0 means that there is only one ethnic group while higher values signify greater levels of ethnic diversity. While diversity itself is not a cause for grievances, the idea with ethnic fractionalization is that the more distinct ethnic groups there are in society, the more competition their is for political power, land, and resources. Ethnic groups further provide a basis for parochialism (in-group love and out-group animosity), which as Blattman (2023) notes can be an added source of antagonism between groups.

As discussed in the previous chapter, the Polity index is one among several measures of democracy. Collier and Hoeffler (2004) argue that poorer democracy scores also imply higher levels of government repression. Such repression can be a basis for righteous outrage against the government, incentivizing violent popular responses.

The Gini index is a popular measure of economic inequality. Higher values imply greater inequality within a society. Collier and Hoeffler (2004) contend that high levels of inequality can be a basis for grievances among an impoverished society against a wealthy elite.

Finally, ethnic dominance captures whether a single ethnic group makes up a majority of the population. If there is one group that can dominate the others, this might further be grounds for grievances on the part of the smaller ethnic groups toward the largest and most powerful.

Each of these measures has some strengths and also some weaknesses, but taken together they help to triangulate the concept of grievance that we really would like to measure. This idea of using proxy factors for a concept of interest is what we’ll do in the next section. We won’t use all the same variables as Collier and Hoeffler (2004), but we’ll do our best to come up with factors that hopefully capture multiple dimensions of grievance.

9.2 Testing Role of Grievance with Data

For our analysis we’ll focus on three different factors that proxy for grievances:

Ethnic fractionalization (Nardulli et al. 2012)
V-Dem democracy scores (Coppedge et al. 2020)
Unemployment rate (Feenstra, Inklaar, and Timmer 2015)

The measure of ethnic fractionalization is such a popular measure in political science research that {peacesciencer} comes with a function for adding this variable to a base dataset. This will be the only measure that we use that is identical to Collier and Hoeffler (2004).

To measure democracy, we’ll use the Varieties of Democracy Project’s measure of quality of democracy. The reason is that, as summarized in the last chapter, V-Dem’s measure emphasizes certain factors that would bring in more in line with capturing variation in grievances. Namely, it emphasizes free and fair elections and freedom of media—factors that would be absent in countries with repressive governments.

To capture economic grievances, we’ll use unemployment, which comes from the Penn World Table version 10.01. The idea is that the higher the unemployment rate, the more economic grievance we might expect to see in society.

We’ll use these measures to explain variation in civil war onset. The reason is that grievances within a country are little easier to proxy with these measures than grievances between countries or between leaders of countries. To measure civil war onset, we’ll use the popular UCDP/PRIO Armed Conflict Database (Gleditsch et al. 2002; Pettersson, Högbladh, and Öberg 2019). This measure will take the value of 1 for a country-year in which a new civil war starts; it will be 0 otherwise.

If grievances are indeed a reason for war, a nature set of hypotheses with our data is that:

\(H_1\): Higher ethnic fractionalization will lead to an increase in the likelihood of civil war onset.
\(H_2\): Higher quality of democracy will lead to a decrease in the likelihood of civil war onset.
\(H_3\): Higher unemployment will lead to an increase in the likelihood of civil war onset.

With our variables selected, the next section walks through the code necessary to construct a dataset for analysis.

9.3 Create the Data

Our base dataset will be at the level of country-years from 1950 to 2013. Our analysis will be restricted to this time span because this is the range of years for which we have full coverage for our variables of interest. The below code opens the {tidyverse} and {peacesciencer} packages. It then use create_stateyears() to create a state-year level dataset from 1950 to 2013 using the Gleditsch-Ward (GW) state system. The GW system is the one that the creates of the Armed Conflict Dataset used in their data collection, so it makes sense to use it has our basis for creating our data.

## packages
library(tidyverse)
library(peacesciencer)

## construct base data of country-years ('50 - '13)
create_stateyears(
  system = "gw",
  subset_years = 1950:2013
) -> Data

The next step is to populate our data with our variables of interest. The below code adds to the data the UCDP/PRIO data, a measure of civil war peace spells, ethnic fractionalization data and democracy data. Notice that in add_ucdp_acd() the option only_wars = T. The ACD dataset uses two different battle deaths thresholds for including conflicts in the data. The first is a 25 battle deaths threshold. ACD codes these conflicts as minor. The second is a 1,000 battle deaths threshold, which ACD codes as all out wars. For our analysis, we’ll focus on all out wars since this raises the bar for what counts as sustained organized conflict versus more random flareups of violence.

## populate data with {peacesciencer} variables for analysis
Data |>
  ## add conflict data
  add_ucdp_acd(only_wars = T) |>
  add_spells() |>
  ## add fractionalization data
  add_creg_fractionalization() |>
  ## add democracy data
  add_democracy() -> Data

Joining with `by = join_by(gwcode, year)`

Warning in xtfrm.data.frame(x): cannot xtfrm data frames

Joining with `by = join_by(orig_order)`
Joining with `by = join_by(gwcode, year)`
Joining with `by = join_by(gwcode, year)`

Unemployment data isn’t available with {peacesciencer}, so we need to pull it from somewhere else. In particular, we’ll access it using the {pwt10]} R package, which lets R users access the latest version of the Penn World Table dataset. This dataset contains a number of economic variables for countries from 1950 to 2019. Most importantly for our analysis, it contains employment data. The below code opens the {pwt10} package. It then selects from the pwt10.01 data object the columns that we need for our analysis. In particular, it pulls out country ISO codes, the year, population, and employment. Before saving this narrowed down version of the data, a new column is added to the dataset that uses employment and population data to construct a measure of the unemployment rate.

## use penn world table data
library(pwt10)

## we want 
## - pop: population
## - emp: employment
pwt10.01 |>
  select(
    isocode,
    year,
    pop,
    emp
  ) |>
  mutate(
    ## make unemployment rate measure
    unemp_rate = 1 - emp / pop
  ) -> lab_Data

We now need to cross-walk the economic data to ensure we align the right rows in it with the correct rows in our main dataset. This will let us then merge the datasets together. The below code uses countrycode() from the {countrycode} package to convert ISO codes to GW codes. It then merges the economic data to the main dataset using a left-join by GW codes and the year.

## cross-walk lab_Data to Data
library(countrycode)
lab_Data |>
  mutate( 
    ## make gwcode column to match Data
    gwcode = countrycode(
      isocode, "iso3c", "gwn"
    )
  ) -> lab_Data

Warning: There was 1 warning in `mutate()`.
ℹ In argument: `gwcode = countrycode(isocode, "iso3c", "gwn")`.
Caused by warning in `countrycode_convert()`:
! Some values were not matched unambiguously: ABW, AIA, ATG, BMU, CUW, CYM, DMA, GRD, HKG, KNA, LCA, MAC, MSR, PSE, STP, SXM, SYC, TCA, VCT, VGB, YEM

## join to Data by gwcode and year
Data |>
  left_join(
    lab_Data,
    by = c("gwcode", "year")
  ) -> Data

9.4 Perform the Analysis

There are many complicated analyses we might consider doing with this dataset. As a first pass, a simple approach is to estimate a regression model of conflict onset using each of the measures of interest. As we’ve done again and again, we’ll also include a cubic peace spells trend in our model. The below code, in particular, estimates a logit model where civil war onset is a function of ethnic fractionalization, quality of democracy, and the unemployment rate.

## estimate the model
glm(
  ucdponset ~ ethfrac + v2x_polyarchy + unemp_rate +
    poly(ucdpspell, 3),
  data = Data,
  family = binomial
) -> model

With our model now fit to the data we can use plot_model() from the {sjPlot} package to visualize the output. So that we can see the output for each measure of interest side-by-side, the below code relies on the {patchwork} package to put separate plots of marginal effects into a single figure. As the output shows, each of the measures operates in the direction we’d expect if grievances were a motivation for war. Both higher ethnic fractionalization and unemployment rates have a positive relationship with the likelihood of civil war onset. Meanwhile, quality of democracy has a negative relationship with the likelihood of civil war onset.

## visualize the output 
library(sjPlot)
library(patchwork)

plot_model(
  model,
  type = "eff",
  terms = "ethfrac"
) +
  labs(
    x = "Ethnic Fractionalization",
    y = NULL,
    title = NULL
  ) +
  scale_x_continuous(
    labels = scales::percent
  ) -> p1
plot_model(
  model,
  type = "eff",
  terms = "v2x_polyarchy"
) +
  labs(
    x = "Democracy (V-Dem)",
    y = NULL,
    title = NULL
  ) -> p2
plot_model(
  model,
  type = "eff",
  terms = "unemp_rate"
) +
  labs(
    x = "Unemployment Rate",
    y = NULL,
    title = NULL
  ) +
  scale_x_continuous(
    labels = scales::percent
  ) -> p3

p1 + p2 + p3 +
  plot_annotation(
    title = "Logit Marginal Effects"
  )

9.5 Conclusion

This analysis doesn’t even scratch the surface of what’s possible in testing intangible incentives, but it is a good start. Encouragingly for Blattman (2023), this output from the above example is consistent with the argument that intangible incentives can be a reason for war. It’s important to note, however, that this analysis is contingent on the idea that the explanatory variables used actually can proxy for grievances. Since we cannot measure grievances directly and in a consistent way across countries, we have to trust that our observable factors are closely correlated with what we really want to measure. This is a common limitation in conflict research. But it also means that there is ample room for a motivated future scholar to do the hard work of devising a better measure of grievances for use in analysis.