Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Analysis of Armed Conflict Data in Africa
Chloe Fearn Supervisor: Christian Rohrbeck
Lancaster University
August 30, 2017
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
1 Introduction
2 Exploratory Data Analysis
3 Modelling Framework
4 Modelling Africa
5 Results
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Introduction
Is there a relationship between level of cropland in Africa andincidents of violence against civilians?
Using data collected by the Armed Conflict & Event DataProject (ACLED) between 1997-2009.
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Data
Data collected on an annual basis in blocks of 0.5◦ longitude× 0.5◦ latitude.
There are 296 covariates.
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Data
The response variable, n, is the number of incidents ofviolence per year in a 0.5◦ × 0.5◦ square.
Used AIC to decide which covariates were best to include inthe model.
The covariates that were selected to be included in the designmatrix, X, are:
cropland,log(population), andlog(GCP).
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Model Fitting
Figure: A map to show croplanddensity in Africa.
Figure: A map to show frequencyof incidents of violence in Africa.
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Hurdle Model
The response variable is 97% zeros so we use a hurdle model.
First we consider P (N > 0) .
Then we consider P (N = n | n > 0) .
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Modelling P (N > 0)
We fit a logistic regression model with
logit [ P(N > 0) ] = β̂̂β̂βTX ,
where β̂̂β̂β is the vector of parameter estimates, and
logit (p) = log
(p
1− p
).
Binary response.
Covariates can be discrete or continuous.
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Modelling P (N = n | n > 0)
We find that the Poisson model is not flexible enough to be agood fit for the data.
Figure: A QQ plot to show the poor fit of the Poisson model to the data.
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Modelling P (N = n | n > 0)
Instead of Poisson, we consider a negative binomial model,and find that the fit is much better.
Figure: A QQ plot to show the fit of the negative binomial model to thedata.
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Koren & Bagozzi
In the study, Living Off The Land: The connection betweencropland, food security, and violence against civilians, Koren& Bagozzi found:
cropland adds a pacifying effect when there is already peace,andcropland increases violence against civilians where there isalready conflict.
However, they did not take into account that some countries’violence may react differently to changes in cropland.
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Modelling Africa
When we fit all of Africa, the fit isnot ideal.
The association between number ofincidents of violence and ourcovariates varies between countriesin Africa.
We can test whether there is adifference in values for β̂̂β̂β using alikelihood ratio test.
Figure: A QQ plot to showthe fit of the negativebinomial model for all ofAfrica.
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Likelihood Ratio Test
Our null and alternative hypotheses are:
H0 : βββ1 = βββ2 = . . . = βββ26 , and
H1 : βββ1 6= βββ2 6= . . . 6= βββ26 .
The test statistic is
D = 2 (l̂ for alternative model− l̂ for null model)
= 2 (26∑k=1
l̂k − l̂ )
which we compare with a critical value from the χ2
distribution.
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Likelihood Ratio Test
When we calculate the critical value and test statistic, we findthat, using the likelihood ratio test, we reject the null.
Negative binomial Logistic regressionCritical value 152 124
Test statistic 846 2376
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Grouping Countries
Since we have found that it is better to consider countriesseparately, we can group them into neighbouring countriesand assess the fit of the models.
Red: Algeria, Morocco.
Green: Senegal, Guinea, Sierra Leone,Liberia, Ghana.
Blue: Eritrea, Ethiopia, Kenya, Uganda,Rwanda, Burundi, Tanzania.
Cyan: Malawi, Mozambique, SouthAfrica, Namibia, Angola, Zambia.
Magenta: Niger, Nigeria, Chad,Cameroon, Central African Republic,Congo.
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Results
Figure: A QQ plot to show the fit ofthe negative binomial model for allof Africa.
Figure: A QQ plot to show thenegative binomial model fit of thegroup of countries in the south.
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Results
We can use AIC to test model fit.
AIC = 2k − 2l̂
where k denotes number of parameters in the model and l̂denotes the log(MLE).
AIC negative binomial AIC logistic regressionSeparate 6252 13679
Combined 6507 14996
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa
Introduction Exploratory Data Analysis Modelling Framework Modelling Africa Results
Thank You
Thank you for listening.Any questions?
Chloe Fearn Supervisor: Christian Rohrbeck Lancaster University
Analysis of Armed Conflict Data in Africa