UNHCR TS Analysis.Rmd

---
title: "UNHCR Time Series Analysis and Asylum Seeker Forecasting"
output:
  pdf_document: default
  word_document: default
---
```{r global_options, include=FALSE}
knitr::opts_chunk$set(error = FALSE, warning = FALSE, cache = TRUE, message=FALSE)
```

#1.1 Prepare Dataset for Analysis
```{r setup, include = FALSE}
library(xml2)
library(dplyr)

#plots
library(devtools)
library(ggplot2)
library(plotly)
library(reshape2)
library(textreg)

#time series
library(dynlm)
library(urca)
library(tseries)

#forecasting
library(forecast)
library(scales)

#knitting
library(knitr)
library(rvest)
```

##1.2 Import Data
Install packages and load libraries, set working directory
```{r results='hide', warning=FALSE, message=FALSE}
setwd("/Users/tessaschneider/Desktop/Final Data Analysis")

df <- read.csv("unhcr_popstats_refugee-status.csv", skip = 2, stringsAsFactors = F, 
               na.string=c("", "*"))
df2 <- read.csv("unhcr_popstats_export_persons_of_concern_all_data.csv", skip = 3, 
                stringsAsFactors = F, na.string=c("", "*"))
```

##1.3 Tidy Data
Convert Total.Population to numeric before merging the datasets (after merging, data before 2000 drops out)
```{r warning=FALSE, message=FALSE}
df2$Total.Population[df2$Total.Population=="*"] <- "2.5"
df2$Total.Population <- as.numeric(df2$Total.Population)

merged_data <- merge(df2, df, by = c("Year", "Country...territory.of.asylum.residence",
                                     "Origin"))

table(merged_data$Year)
```

##2.1 Exploring Data: Top Destination Countries

```{r, warning=FALSE, message=FALSE}
destination_country_total <- merged_data %>%
  group_by(Country...territory.of.asylum.residence, Year) %>%
  summarise(Total = sum(Total.Population))

top_destcountries <- destination_country_total %>%
  group_by(Country...territory.of.asylum.residence) %>%
  summarise(Total = sum(Total, na.rm = TRUE)) %>%
  top_n(20)

top_destcountries2 <- as.character(top_destcountries$Country...territory.of.asylum.residence)

destination_country_total %>%
  filter(Country...territory.of.asylum.residence %in% top_destcountries2) %>%
  ggplot(mapping = aes(x = Year, y = Total)) +
  geom_line() + coord_cartesian(ylim = c(0, 3e6)) +
  facet_wrap( ~ Country...territory.of.asylum.residence, ncol=4)

```

##2.2 Exploring Data: Top Origin Countries

```{r warning=FALSE, message=FALSE}
origin_country_total <- merged_data %>%
  group_by(Origin, Year) %>%
  summarise(Total = sum(Total.Population))
  
top_origcountries <- origin_country_total %>%
  group_by(Origin) %>%
  summarise(Total = sum(Total, na.rm = TRUE)) %>%
  top_n(20)

top_origcountries2 <- as.character(top_origcountries$Origin)

origin_country_total %>%
  filter(Origin %in% top_origcountries2) %>%
  ggplot(mapping = aes(x = Year, y = Total)) +
  geom_line() + coord_cartesian(ylim = c(0, 1e7)) +
  facet_wrap( ~ Origin, ncol=4)

table(merged_data$Year)
```

##2.3 Exploring Data: Percent Change in Total Population
By "People of Concern"", subset for only PoC category counts by year
change value from character to integer
```{r warning=FALSE, message=FALSE}
Year_Pop <- aggregate(merged_data$`Total.Population`, by=list(Year = merged_data$Year), 
                      FUN=sum, na.rm = TRUE)

Year_Pop$rate <- NA

Year_Pop$rate[which(Year_Pop$Year>2000)] = 100*(diff(Year_Pop$x)/Year_Pop[-nrow(Year_Pop),]$x)

ggplot(Year_Pop, aes(x= Year, y= rate)) + geom_line() + 
  labs(title="Percent Change in People of Concern",
       subtitle="(2000 - 2016)",
       x="Year", 
       y="Percent Change")

PoC_count <- merged_data[c(1,4:10)]

PoC_count <- melt(PoC_count, id=c("Year"))

str(PoC_count)

PoC_count$value <- as.integer(PoC_count$value)
```
Starting from 2013 the number of refugees has increased dramatically and with it pending cases for asylum seekers have also increased
```{r}
ggplot(PoC_count,aes(Year,value, na.rm = TRUE)) +
  geom_bar(aes(fill=variable),stat="identity") +
  labs(title="UNHCR Population Statistics Database",
       subtitle="(2000 - 2016)",
       x="Year", 
       y="Number of People (Millions)")
```

##3.1 Time Series Analysis: Preparation
* y is PoC in Germany
* x is PoC in all countries in database
* t is Years (2000-2016)

All variables used in the model must be declared as time series
```{r warning=FALSE, message=FALSE}
Germany_PoC <- merged_data %>% group_by(Country...territory.of.asylum.residence, Year) %>% 
  filter('Germany' %in% Country...territory.of.asylum.residence) %>% 
  summarise(Total = sum(Total.Population, na.rm = TRUE))

Germany_data <- merge(Germany_PoC, Year_Pop, by = "Year")

Germany_data$Year <- ts(Germany_data$Year)
Germany_data$Total <- ts(Germany_data$Total)
Germany_data$x <- ts(Germany_data$x)
```

##3.2 Time Series Analysis: Test for Time Series Problems
###Test for Persistence or Dependence
Row is <1 so it meets the stability condition for weak dependency
```{r warning=FALSE, message=FALSE}
summary(dynlm(Total ~ L(Total, 1), data = Germany_data))
```
###Test for Persistence or Dependence
Germany's Total persons of concern annual data shows that the correlation of lags of the Total Population variable drops to zero after 1 lag with statistical insignificant correlation after 1 lag, therefore it is not persistent
```{r}
acf(Germany_data$Total, na.action = na.pass, lag.max = 5)
```

###Tests for Stationarity
Germany Total PoC annual is trending after 2012
Stochastic trend (increases and decreases inconsistently) in the Germany Total plot Deterministic trend (increases and decreases consistently) in the Germany x plot
```{r}
ggplot(data=Germany_data,
       mapping = aes(x = Year, y = Total)) + geom_line()

par(mfrow = c(1,2))
plot(Germany_data$Total)
plot(Germany_data$x)
```

###Tests for Stationarity - Unit Root Test - Dickey Fuller Test 
(p value <.05 then there is no unit root)
```{r}
adf.test(Germany_data$Total)
```

###Detrend: When there is a Deterministic Trend
Regress y, x1 and x2 on trend term(Year) and intercept, save residuals for y, x1 and x2, and then regress y residual on x1 residual and x2 residual
The regression with residuals shows an increase in the correlation, but it is still not statistically significant
Even after detrending there is still no statistically significant coefficient
```{r}
fit = lm(Germany_data$Total ~ Germany_data$Year, na.action = NULL)
plot(resid(fit), type="o", main="Detrended")

fit1 <- lm(Germany_data$Total ~ Germany_data$Year)
res_Germany_dataTotal <- residuals(fit1)

fit2 <- lm(Germany_data$x ~ Germany_data$Year)
res_Germany_datax <- residuals(fit2)

summary(m3 <- dynlm(res_Germany_dataTotal ~ res_Germany_datax))
```

###Detrend: When there is a Stochastic Trend
First differencing then plotting shows that the trend was removed in this case
```{r}
diff_Germany_dataTotal <- c(NA, diff(Germany_data$Total))
diff_Germany_datax <- c(NA, diff(Germany_data$x))

par(mfrow = c(1,2))
plot(diff_Germany_dataTotal)
plot(diff_Germany_datax)
```

##3.3 Run OLS regression
This time series regression resulted in no statistically significant correlation between the selected variables
Since there is monthly data on asylum seekers, perhaps it is possible to predict future numbers of asylum seekers in Germany through a forecasting model
(there are clear limitations in only looking at one variable, so these predictions cannot be interpreted as exact predictions)
```{r warning=FALSE, message=FALSE}
summary(m1 <- dynlm(Germany_data$Total ~ Germany_data$x, Germany_data$Year))
summary(m2 <- dynlm(diff_Germany_dataTotal ~ diff_Germany_datax))
```

##4.1 Forecasting Number of Future Asylum Seekers in Germany: Preparation
As before, we convert values to numeric, create an object that sums all origin countries to Germany by month, declare variables as time series variables
```{r warning=FALSE, message=FALSE}
df3 <- read.csv("unhcr_popstats_export_asylum_seekers_monthly_2017_12_04_203715.csv", 
                skip = 2, stringsAsFactors = F)

df3$Value[df3$Value=="*"] <- "0"
df3$Value <- as.numeric(df3$Value)

Germany_monthlyasylum_total <- df3 %>%
  group_by(Country...territory.of.asylum.residence, Year, Month) %>%
  summarise(Total = sum(Value))

Germany_monthly <- ts(Germany_monthlyasylum_total$Total, 
                      start = c(1999, 1), frequency = 12)
```
##4.2 Forecasting Number of Future Asylum Seekers in Germany: Test for Time Series Problems
###Stationarity Test
Plot and observe trends
```{r}
autoplot(as.zoo(Germany_monthly), geom = "line")
```

###Persistence Test 1
After dynlm, row is <1 so it meets the stability condition for weak dependency)
```{r}
summary(dynlm(Germany_monthly ~ L(Germany_monthly, 1)))
```

###Persistence Test 2
After acf, Germany monthly's correlation of lags drops to zero after 2.5 lags therefore it is not persistent
```{r}
acf(Germany_monthly, na.action = na.pass, lag.max = 40)
```

###Persistence Test 3
After Dickey Fuller Test for Unit Root, p value is <.05 then there is no unit root)
```{r}
adf.test(Germany_monthly)
```

##4.3 Forecasting Number of Future Asylum Seekers in Germany
###Decompose
Then we can decompose the additives of time series.  This returns estimates of the seasonal component, trend component and irregular components or "random"
```{r}
plot(decompose(Germany_monthly))
```

##4.4 Forecasting Number of Future Asylum Seekers in Germany
###Seasonal Changes
To look more closely at the seasonal changes in the number of asylum seekers we use the "stl" function
Germany has had a positive net flow of asylum seekers in June, July, November, December and the highest typically in February between 2000 and 2015
```{r, eval = FALSE, include = FALSE}
stl(Germany_monthly, s.window="periodic")
```

##4.5 Forecasting Number of Future Asylum Seekers in Germany
The ARIMA forecasting method shows possible future changes in the number of asylum seekers in Germany in the next years
The wide confidence intervals show the uncertainty in forecasting with the dark grey representing 95 percent confidence and the light grey representing 80 percent confidence
```{r}
plot(forecast(auto.arima(Germany_monthly), 30), 
     main = "ARIMA Forecast: Germany Asylum Seeker Arrivals", 
     ylab = "Number of Asylum Seekers", 
     xlab = "Year", ylim=c(0, 90000))
```

###ARIMA Forecast Values
```{r}
forecast(auto.arima(Germany_monthly), 24)
```

The TBATS forecasting method shows another possible future change in the number of asylum seekers in Germany in the next years
```{r}
plot(forecast(tbats(Germany_monthly), 30), 
     main = "TBATS Forecast: Germany Asylum Seeker Arrivals", 
     ylab = "Number of Asylum Seekers", 
     xlab = "Year", ylim=c(0, 90000))
```

###TBATS Forecast Values
```{r}
forecast(tbats(Germany_monthly), 24)
```