The regression they are running is whether or not you have a civil war in a particular country, for a particular year, and what was the per capita income, the extent of primary exports, and the extent of ethnic diversity of the population. To what extent are these factors predictors for the likelihood of civil war. Correlations not causal as there is reverse causality and massive feedback effects.

They use a probit and a tobit regression, as these are the regressions used when the outcome variable is binary, when they are either zero or one.

• They are like twins

• A tobit is used for the duration, for how long this happens

• A probit is whether or not it happens or not

Based on these regressions there are some takeaways:

• All of these are correlations

• Higher the per capita income, the lower the risk of civil war. This is the negative coefficient on income, and the t-ratio is above 2, so it is quite significant. The duration is also above 2. So, richer countries typically do not experience too many civil wars and if they do the duration is also very short.

• The effect of natural resources is non-monotonic, the primary commodity export is the measure of natural resources and is non-monotonic as they included both the linear term and the square term (quadratic). One is positive and one is negative which means that initially as you increase natural resources the chance of civil war goes up, but beyond a certain threshold the chances of civil war actually go down. If you have a lot of natural resources then the government can extract a lot of revenue and support a very strong military, "if you are rich enough you are safe".

• On the linear ethnic-diversity variable, political scientists have been saying if a country is very ethnically diverse it is going to have a larger likelihood of civil war. This is not picked up in these correlations, as they are not significant. So it will not be more likely of a civil war if they are more ethnically diverse.

• On the quadratic ethnic-diversity variable, conditional on the civil war breaking out it is likely to be drawn out for longer if there is more diversity. Again, with the quadratic mechanism it will go up and then if diversity is high then the duration is not that long. From absolute homogeneous to completely diverse it increases the duration of a civil war should it happen, but once it becomes very diverse the duration actually falls.

• The ethnic diversity variable and income per capita relations are the most important correlation takeaways. 

What is the maximum ethnic diversity you can have in a country?

• Theoretically, the maximum diversity occurs when every individual belongs to a distinct ethnic group.

• To have a sense of group identity you need to have enough people who identify with your group values, the way of life, cultural, economic interests and so on. You need a minimum amount of closeness to have a team. If the country is absolutely homogeneity you'll likely have other factors which result in differences between them, so you need to be in the sweet spot.

• Both economists and political scientists have postulated that ethno-linguistic fractionalisation is unambigously conflict-enhancing.

• If you have a country with two ethnic groups, and they are similarly sized, that is where the potential for civil war is the highest as each group would be inclined to control the government. The peak is when the index is such that you have two similarly sized groups, and that is not very ethnically diverse.

Economic Shocks & Civil Wars

"Economic Shocks and Civil Conflict: an Instrumental Variables Approach" - E. Miguel; S. Satyannath, E. Sergenti (2004)

• Similar topic but tries to get at causal effects

• This paper focuses on the negative correlation that the previous research highlights the association between economic conditions/performance and civil conflict [review in Sambanis (2001)]

• If a country has lower GDP per capita they are more likely to be engaged in civil conflict. Endogeneity of economic variables to civil war cannot establish causality

• However, we economists want to know if there is in fact a casual effect and if there were negative shocks to your GDP per capita, would this increase the country's likelihood of civil war or not.

• This becomes an experiment where you shock the economy, and see if there is a spurt in the probability of civil war.

Same with Acemoglu, Johnson and Robinson, we know that good institutions foster economic growth, but they hadn't established causality so they used that instrument using colonisation with the disease environment.

This paper uses exogenous variation in rainfall as an independent variable for income growth to estimate the impact of economic growth on civil conflict.

Their sample consists of Sub-Saharan African countries:

• Countries with on average low GDP per capita

• Structural transformation has been so that it is still agricultural which is a primary sector in terms of employment and GDP (value added).

• In most of these countries, most of the agriculture is rain-fed.

• Plausible strategy for economies that rely on rain-fed agriculture: as only 1% of cropland is irrigated in the median African country, and the agricultural sector is large (World Development Indicator database).

• Therefore, if the rains are good you get a good harvest, agriculture sector booms and GDP per capita is now relatively high. But if they experience a drought, which occurs frequently here, then they experience a negative shock through the agricultural sector and GDP per capita tanks. This is where we get the variation from. 

• Therefore, the instrument is exogenous variation in rainfall as it is correlated with economic growth as poor (good) harvest from a drought results in lower (higher) GDP per capita.

• The second requirement of an instrument is that rainfall effects the onset of civil conflict only through the channel of economic growth and nothing else, and that is the exclusion restriction.

• There are of course other channels, but we can address that later...

• A lot of economists say that if you are looking at causal effects in a cross-sectional analysis of countries then the variation is actually is very difficult to make sense. So, you look at variation within the country.

E. Miguel (2005) - "Poverty & Witch Killing"

• Utilises rainfall shocks as an instrument and was subsequently called the Rain-man among peers. 

• After the previous article and this being published in top journals, many researchers thought that they may have solved the problem for a good instrument which was rainfall shocks. 

• In this context it works because you are looking at majority rain-fed agriculture.

• If you were at how rainfall shocks on rain-fed agriculture being an instrumental variable for economic growth in the UK, you would likely not find any correlation as agriculture relies on irrigation in the UK, and agriculture isn't the dominant sector in the UK either.

• So rainfall shocks on economic growth is very context specific.

Data and Measurement

Data on Civil Conflict

They use data on civil war which is much more detailed, one which is open and free and in the form of tables, titled Armed Conflict Data. The database was developed by the International Peace Research Institute of Oslo, Norway, and the University of Uppsala, Sweden (referred to as PRIO/Uppsala).

An armed conflict is defined as a "contested incompatibility which concerns government and/or territory where the use of armed force between two parties, of which at least one is the government of a state, results in at least 25 battle-related deaths (low-intensity conflict).

The database is careful to focus only on politically motivated violence. There are other types of organised violence such as crime, unrelated to the conflict (smuggling rackets, importing rackets).So, it does not capture many important types of organised violence in sub-Saharan Africa that do not directly involve the state.

There are also cases in ethnic groups engaging against each other as it is of course very diverse in ethno-linguistic groups so it doesn't focus specifically on that, it focuses on civil wars.

Therefore, the authors focus exclusively on civil wars. 

Data on Rainfall

They use the database of Global Precipitation Climatology Project (GPCP) of monthly rainfall estimates which stretch back to 1979.

The rainfall estimates are at different nodes for each point at which latitude and longitude degree lines cross, at 2.5 degree intervals. So it is not country based:

• A country such as Kenya, a medium-sized African country contains eight rainfall data "nodes".

• Whereas the largest country, Sudan, contains 34 nodes.

The GPCP rainfall measure at latitude-longitude degree node point p in country i during month m of year t is denoted Ripmt. 

For that country they take an average of all those nodes to create the average annual rainfall for that country, as this analysis is done at the country year level as civil wars don't change month to month. The data rainfall is quite rich, many of the other variables you only get annual data.

Denote the average rainfall across all points p and months m for that year as Rit.

The principal measure of a rainfall shock is the proportional change in rainfall from the previous year 

denoted ΔRit

The idea being that usually there are droughts, so if there is more rain than before, that should impact the output in a positive way.

Data on Other Country Characteristics

The remaining data is drawn from Fearon and Laitin (2003) and from World Bank databases. 

The main two control variables they account for are of course ethno-linguistic fractionalisation (drawn from the Soviet Ethnographic index, Atlas Marodov Mira) and religious fractionalisation (based on the CIA Factbook).

As the ethno-linguistic component was found to be correlated with civil war and that it could be correlated with economic growth, you need to control for that, as if you don't you have omitted variable bias.

Measures of democracy (from the Polity IV data set as it gives a rank), the log of per capita income (from the Penn World Tables and the World Bank), the proportion of a country that is mountainous, as this has impact on agricultural output.

They have been careful not to include:

• Income inequality

• Poverty

• Measures of unemployment rates

as additional explanatory variables because of the large number of missing or unreliable African observations in existing macroeconomic series. These are potentially omitted variables.

You could argue that rainfall is purely exogenous.

The first bit of data is on conflict, Panel A, Panel B is rainfall, Panel C is economic growth, 

So you want the effect of Panel A on Panel C, and Panel B is your instrumental variable. Panel D is a bunch of country characteristics that they control due to omitted variables. Interested in the causal effect of economic growth on civil conflict, does economic growth dampen the chances of civil war? Economic growth has a two way relationship (if there is civil war then naturally there will be lower economic growth).

The Variables

They have evidently tried a whole bunch of different conflict variables. 

1. For the Classification with 25 annual battle related deaths.

• The mean is .27 and number of observations is 743. Looking at the span of years of 1981-99 there are 18 years , and if they have 743 observations, that means 743/18 is the number of countries that they have.

• They are tracking that many countries over this period

• The value 0.27 means that for a given year this variable for a given country either takes a value of zero or one, was there civil conflict in that year for that country which had at least 25 battle related deaths. The mean will inherently be between zero and one, being 0.27 which means in 27% of the 743 cases there was actually civil war conflict going on. 

• So, for more than a quarter of the time in that period there was some conflict going on.

2. They can break it up by onset when it started, or the offset of when it closed.

3. The stricter definition of 1,000 deaths is a jump which is significant but it doesn't disappear at the mean is still 0.17 so it is not likely you are loosing out on a lot, but still something. In 17% of the cases there were people who were dying, and this is still undercounting.

4. The average rainfall in ml is 1,000 with a lot of variation, the standard deviation is 500 over time and across the country.

• Rainfall growth which is their main variable is also computed, and we get a mean and standard variation.

• For economic growth, it is actually negative in that period. We are really talking about sub-saharan Africa which aligns with what we find. They countries were facing negative shocks.

5. GDP per capita

6. Democracy indicators

7. Extent of ethnic fractionalisation (a probability that the two people randomly picked up are from different groups), and the probability is 0.65 in the sample, which indicates heterogeneity. So, if they pick up two people there is a 65% chance that they are from different groups. 

• When using the CIA stuff it becomes 0.49.

Important Note: If the descriptive statistics do not look right, go back to the data and check if it is okay.

Estimation Framework

Instrumental Variables - Two-Stage Least Squares Approach

An instrumental variable is required, a simple OLS will not get you the correct measure because of this reverse causality. You must account for endogeneity by using this instrument:

• An instrument is a variable used when inspecting how x affects y, where you introduce this instrumental variable as z with these characteristics:

  • z is correlated with x (something strongly correlated with economic growth)

  • Exclusion restriction: z must not be directly affecting y by itself

• Rainfall shocks are introduced as the instrumental variable, rainfall shocks are potentially related with economic growth in the sample  of countries being examined in sub-saharan Africa primarily reliant on rainfall fed agriculture. Agriculture is the dominant sector (main growth component of GDP) and a positive rainfall shock leads to higher amount of output through better harvests and therefore more harvest.

• Annual rainfall shocks do not directly affect the likelihood of conflict, except for its effect through economic growth channel.

An instrumental variables approach is often consider a two staged least squares approach as you look at the instrument and the endogenous variable. The first stage relation is between rainfall shocks and growth, as if this is not true then the instrument is invalid as there is no knowing when using the instrument what you are picking up.

First-Stage Relationship

where

• ai = Country Fixed Effects are included in some specifications to
  capture time-invariant country characteristics that may be related to civil conflict.

  • Whenever using any variable which is such like "what proportion of the country is mountainous" you can't use this, as this would become perfectly correlated.

• Also include country-specific time trends in most specifications to capture additional variation. Our t are our year dummies to account for any annual trends, these are shocks which affect all of the countries in the sample in a given year, you de-trend this as you want to inspect the specific rainfall effect in that country.

• The term e is a disturbance term, and these disturbances are allowed to be correlated across years for the same country in all regressions. Captures everything which has not been accounted for.

Growth in country i over a period of time t is dependent on a multitude of factors, but most importantly the rainfall shock variables (ΔRit and ΔRi,t-1), growth in rainfall over previous years. Essentially the annual growth in rainfall.

• The authors include not just one-period of growth but also the lag of that, because of course one year of bad rainfall with effect crops and harvests which may persist for longer than a year or two.

Regressions like these, where you observe a certain unit over a number of years, and you have a bunch of such units, is referred to as Panel Data Regressions.

Below are the results from the effect of rainfall shocks on economic growth.

First Column:

• Growth in rainfall in that year (Growth in rainfall t), and the previous year are both positive and statistically significant, implying that if you had a good rainfall in this year and the previous year, economic growth in the current year is high.

  • We find evidence of our relation, but more robustness is done to add more variables.

Second Column:

• Added more variables for robustness, such as:

  • Log GDP per capita

  • level of democracy in the country

  • degree of ethno-linguistic diversity

  • degree of religious diversity 

  • whether it is an oil exporting country

  • what proportion of the country is mountainous

  • population

• After adding all these, the coefficient on rainfall growth and in the current year and the past year is still positive and significant.

• Implying that the instrument is indeed correlated with economic growth.

Column's 3, 4 and 5 continue by adding and removing more control variables to highlight the robustness of the relation between growth and rainfall, and economic growth.

Column 4:

Has three growth in rainfall variables:

• Growth in rainfall in year t (current period)

• Growth in rainfall in year t-1 (last year)

• Growth in the rainfall in year t+1 (the future)

In the first two, the coefficient is positive and significant as rainfall shocks effect the economic growth. The reason for this is referred to as a falsification test as if the third variable was also significant then we would be worried why future rainfall could come back and effect economic growth today.

First Stage Results

The first stage relationship between rainfall and income growth is strongly positive: both the current and lagged rainfall are significantly related to economic growth.

This makes sense as the countries we are investigating as positive rainfall typically leads to better agricultural production since most of sub-Saharan Africa les within the semi-arid tropics and is prone to drought. It is the lack of rainfall which is typically the problem.

Somewhat of a technical point: we need to correlation between the instrument and the endogenous variable to be sufficiently strong. We use the F-statistic which (F-Statistic is 4.5 in regression 3), which is not very high and suggesting that the IV-2SLS estimates may be somewhat biased toward OLS estimates. If it is 10 and above, we are typically happy. 

To bolster the argument that there is a relationship between rainfall shocks and economic growth we do the column four regression. This is the identification check for the "false experiment" specification in which future rainfall shocks on economic growth should be orthogonal to current economic growth, conditional on country-specific time trends is included as an explanatory variable. We find that it is not significant.

The first test has passed, so we now run the main idea regression, economic growth and civil conflict. We also report the OLS results:

• Column 1 uses a probit, as you're investigating whether there was conflict in the year or not. 

• Column's 2, 3, 4 use OLS

• Columns, 5, 6, 7 use rainfall shocks instrumental variable and report the instrumental variable result. Therefore, these are the main results, but the other are still relevant.

Column's 1-4, looking at economic growth coefficient at either the current year or the previous year that does not seem to be correlated with civil war. The coefficients are negative but not statistically significant.

• If they believed it was causal through OLS, they would think that it doesn't matter if the country has more economic growth, it wouldn't lead to civil war. Which obviously is not very accurate.

• This OLS suffers from previous limitations of the previous paper through reverse causation, etc, etc. 

In Column 5, we introduce the rainfall shock instrument and we see the coefficient on economic growth experience an increase in magnitude and become statistically significant. OLS estimates are biased downwards which mean that the actual effect of economic growth is even larger than what was captured in columns 1-4.

In column 5, we have introduce the country characteristics.

In column 6, we introduce the country dummies, and can no longer introduce country characteristics as these are perfectly correlated, and the effects are largely similar. The coefficient on economic growth in the previous year is negative and significant.

A point of concern arises as we expected economic growth in the current year to be negative and significant as well, looking at economic growth rate in year t, is negative but not statistically significant. The growth variable kicks in with a lag.

• If you had a negative economic shock last year, then this year you are more likely to have a civil war, and is not instantaneous. 

The authors did a joint test to see whether you can reject that both are negative and significant, you cannot reject that as growth in the past year and growth in the current year might be correlated to some extent so the negative sign and the significance is not just capturing the previous, it is probably also picking up the current year. So, no need to be concerned. There probably is a time gap in the dynamics.

In column 7, they employ the strict definition of civil war that you need to have at least 1,000 deaths, and the that contemporaneous growth coefficient is negative.

• We notice that the effect is big

• The size of the estimated impact is huge

Focusing on the IV-2SLS fixed-effects specification as the benchmark.

• The point estimate indicates that a 1% point decline in GDP increases the likelihood of civil war conflict by over 2% points.

• A 5% point drop in annual growth increases the likelihood of a civil conflict in the following year by over 12% points which amounts to an increase of more than one-half in the likelihood of civil war.

• IV-2SLS estimate is in fact much larger and more negative than the OLS estimates, which suggests bias due measurement error in the per-capita income growth measures likely to be larger in magnitude than the endogeneity bias, which is presumably negative.

There is an unexpected finding of positive (though insignificant) point estimates on lagged growth in certain OLS specifications (regressions 3 and 4), which casts some doubt on this attenuation bias explanation for the difference between the OLS and IV estimates.

The main contribution lies in the contribution through an instrumental variables for which you can actually get to a causal estimate. 

We have established that in the sample of countries that we have studied, relevant sample as civil wars do happen in more than a quarter of the cases/observation, an increase in economic growth decreases the chances of civil war.