Does hunger make Africans want to move abroad?

Analysis of the relationship between hunger and emigration

knitr::opts_chunk$set(
  out.width = "100%", 
  fig.align = "center", 
  fig.showtext = TRUE
)

Introduction: The Impact of Hunger on African Emigration

The relationship between food scarcity and the desire to leave one’s home country is a complex phenomenon across the African continent. Understanding whether hunger acts as a primary “push factor” for migration requires a deep dive into socioeconomic data and statistical modeling.

Using extensive survey data from Afrobarometer, we analyzed how experiencing hunger affects the likelihood of citizens considering emigration. To quantify this, we utilized Logit Regression models, adjusting for individual-level variables such as age, gender, and education. The core of our analysis relies on Odds Ratios (OR):

• An OR > 1 indicates that hunger increases the likelihood of wanting to emigrate.

• An OR < 1 suggests that hunger actually correlates with a lower desire to move.

In the following chart, we visualize these results across multiple African nations. Each point represents a country, where its vertical position indicates the strength of the relationship (Odds Ratio) and its horizontal position represents Statistical Certainty (P-value). Note that the horizontal axis is plotted on a logarithmic scale to better distinguish between highly significant results (close to 0 on the left) and those that lack statistical confidence (on the right).

Furthermore, we have incorporated a third dimension: Remittances. By color-coding countries based on the percentage of GDP they receive from citizens abroad, we can explore if a national culture of migration influences the individual’s response to food insecurity.

Original graph

Replica

First, in order to create our graph, we load both the libraries needed.

library(tidyverse)
library(scales)
library(ggforce)
library(ggpattern)
library(shadowtext)
library(ggtext)
library(haven)
library(broom)
library(logistf)
library(survey) 
library(grid)
library(patchwork)

In this stage, we import the results generated from the logit regression models. These data points contain the estimated coefficients for each country, specifically the Odds Ratios and their corresponding P-values. This step is crucial as it bridges the raw survey data from Afrobarometer with the statistical evidence needed to visualize the relationship between food insecurity and migration intentions.

source("results.R")

Creating the color of remittances

To explore the influence of external financial flows on migration intent, we categorize countries based on their Remittance-to-GDP ratio. This process involves creating discrete intervals that allow us to map complex numerical data onto a clear visual scale.

By defining specific color palettes for both the fill and border of our data points, we ensure that the final visualization is not only aesthetically pleasing but also functionally intuitive. The gradient reflects the intensity of remittance dependency, helping to highlight patterns between economic support and food-related migration drivers.

# Defining hierarchical levels for remittance groups
remittance_groups_levels <- c("> 40%", "35-40%", "30-35%", "25-30%", "20-25%", 
                              "15-20%", "10-15%", "< 10%")

# Transforming raw data into ordered categorical groups
data_graph <- full_results %>%
  mutate(
    Remittance_Group = case_when(
      Remittances > 40 ~ "> 40%",
      Remittances > 35 & Remittances <= 40 ~ "35-40%",
      Remittances > 30 & Remittances <= 35 ~ "30-35%",
      Remittances > 25 & Remittances <= 30 ~ "25-30%",
      Remittances > 20 & Remittances <= 25 ~ "20-25%",
      Remittances > 15 & Remittances <= 20 ~ "15-20%",
      Remittances > 10 & Remittances <= 15 ~ "10-15%",
      Remittances <= 10 ~ "< 10%",
      TRUE ~ "NA"),
    Remittance_Group = factor(Remittance_Group, levels = remittance_groups_levels))
    
# Color palette for point filling (Fill)
remittance_colors <- c(
  "> 40%" = "#7D6892",
  "35-40%" = "#906386",
  "30-35%" = "#AB6D84",
  "25-30%" = "#C87982",
  "20-25%" = "#E48281",
  "15-20%" = "#F88E7F",
  "10-15%" = "#FFAB9F",
  "< 10%" = "#FDCAC3")

# Color palette for point outlines (Border)
border_colors <- c(
  "> 40%" = "#421C5E",
  "35-40%" = "#6B2E5D",
  "30-35%" = "#8C3757",
  "25-30%" = "#B54B58",
  "20-25%" = "#DA5856",
  "15-20%" = "#F66652",
  "10-15%" = "#FF8978",
  "< 10%" = "#FCB6AD")

Refining Data Coordinates for Visual Fidelity

To ensure our visualization accurately mirrors the original chart’s layout and maintain high legibility, we perform a targeted data manipulation. This step involves micro-adjusting specific P-values and Odds Ratios for several countries.

These manual adjustments are necessary to prevent overlapping data points and to account for the specific logarithmic scaling and coordinate mapping used in the original Afrobarometer publication. By fine-tuning these positions, we enhance the “cleanliness” of the scatter plot without compromising the underlying statistical relationships being depicted.

# Fine-tuning point locations to match the original chart's layout
data_manipulated <- data_graph %>%
  mutate(
    P_Value = case_when(
      Country == "São Tomé and Príncipe" ~ 0.0032,
      Country == "Lesotho" ~ 0.0017, 
      Country == "Nigeria" ~ 0.024,
      Country == "Liberia" ~ 0.017,
      Country == "Malawi" ~ 0.017,
      Country == "Benin" ~ 0.015,
      Country == "Tunisia" ~ 0.027,
      Country == "Mali" ~ 0.035,
      Country == "Côte d'Ivoire" ~ 0.06,
      Country == "Ghana" ~ 0.069,
      Country == "Uganda" ~ 0.13,
      Country == "Namibia" ~ 0.179,
      Country == "Botswana" ~ 0.117,
      Country == "Morocco" ~ 0.61,
      Country == "Niger" ~ 0.33,
      Country == "Cabo Verde" ~ 0.225,
      Country == "Zimbabwe" ~ 0.2,
      Country == "Mozambique" ~ 0.165,
      Country == "Gabon" ~ 0.273,
      Country == "Sierra Leone" ~ 0.28,
      Country == "Zambia" ~ 0.231,
      Country == "Burkina Faso" ~ 0.24,
      Country == "Urban Nigeria" ~ 0.01,
      Country == "Madagascar" ~ 0.8,
      Country == "Senegal" ~ 0.75,
      Country == "Rural Nigeria" ~ 0.4,
      Country == "Cameroon" ~ 0.53,
      Country == "South Africa" ~ 0.65,
      Country == "eSwatini" ~ 0.705,
      Country == "Sudan" ~ 0.65,
      Country == "Kenya" ~ 0.65,
      TRUE ~ P_Value
    ),
    OR = case_when(
      Country == "Nigeria" ~ 0.71,
      Country == "Gabon" ~ 0.85,
      Country == "Zambia" ~ 0.795,
      Country == "Sierra Leone" ~ 0.91,
      Country == "Benin" ~ 1.53,
      Country == "Liberia" ~ 1.35,
      Country == "Tunisia" ~ 1.56,
      Country == "Madagascar" ~ 1.06,
      Country == "Burkina Faso" ~ 0.76,
      Country == "Namibia" ~ 0.71,
      Country == "Rural Nigeria" ~ 0.817,
      TRUE ~ OR
    )
  )

Customizing Label Placement and Alignment

Standard automated labeling in R often leads to overlaps, especially when dealing with a high density of data points on a logarithmic scale. To achieve a professional finish that mirrors the original publication, we implement a manual coordinate calibration for each country name.

In this step, we adjust the x and y positions of the labels and introduce multi-line string formatting (using \n) for countries with longer names. Furthermore, we define custom horizontal and vertical alignments (h_align and v_align) to ensure that text is perfectly positioned relative to its corresponding data point, maintaining readability even in the most crowded areas of the chart.

# Creating manual label offsets
label_positions <- data_manipulated %>% 
  select(Country, P_Value, OR) %>% 
  mutate(label_x = P_Value, label_y = OR) %>%
  mutate(
    label_x = case_when(
      Country == "Nigeria" ~ 0.0123, Country == "Tanzania" ~ 0.0017,           
      Country == "Gambia" ~ 0.00074, Country == "Togo" ~ 0.00135,    
      Country == "Lesotho" ~ 0.00235, Country == "São Tomé and Príncipe" ~ 0.00281,   
      Country == "Benin" ~ 0.011, Country == "Côte d'Ivoire" ~ 0.104,
      Country == "Tunisia" ~ 0.038, Country == "Liberia" ~ 0.0125,
      Country == "Mali" ~ 0.026, Country == "Malawi" ~ 0.0115,
      Country == "Botswana" ~ 0.08, Country == "Mozambique" ~ 0.094,
      Country == "Zimbabwe" ~ 0.132, Country == "Niger" ~ 0.25,
      Country == "Morocco" ~ 0.98, Country == "Madagascar" ~ 1.35,
      Country == "Senegal" ~ 0.51, Country == "Sudan" ~ 0.45,
      Country == "Cameroon" ~ 0.34, Country == "Sierra Leone" ~ 0.175,   
      Country == "Gabon" ~ 0.2039, Country == "Zambia" ~ 0.16,
      Country == "Ghana" ~ 0.069, Country == "Uganda" ~ 0.083,
      Country == "Burkina Faso" ~ 0.295, Country == "Namibia" ~ 0.238,
      Country == "Kenya" ~ 1.05, Country == "South Africa" ~ 1.2,
      Country == "Guinea" ~ 1.08, Country == "eSwatini" ~ 1.1,
      Country == "Cabo Verde" ~ 0.322,      
      TRUE ~ label_x
    ),
    label_y = case_when(
      Country == "Nigeria" ~ 0.71, Country == "Tanzania" ~ 0.528,    
      Country == "Gambia" ~ 1.676, Country == "Togo" ~ 1.846,    
      Country == "Lesotho" ~ 1.648, Country == "São Tomé and Príncipe" ~ 1.537, 
      Country == "Niger" ~ 1.32, Country == "Zimbabwe" ~ 1.163,
      Country == "Sudan" ~ 0.987, Country == "Ghana" ~ 0.84,
      Country == "Burkina Faso" ~ 0.743, Country == "Namibia" ~ 0.695,
      Country == "Cameroon" ~ 0.95, Country == "Zambia" ~ 0.8,
      Country == "eSwatini" ~ 0.924, Country == "Cabo Verde" ~ 1.19,      
      TRUE ~ label_y
    )
  )

Typography and Visual Hierarchy

In this section, we define the typographic styles for the country labels to create a clear visual hierarchy. The styling is not merely aesthetic but serves as a secondary layer of data encoding:

• Sample Size Representation: Countries with larger respondent pools (above 2,200) are assigned a bold-italic style, while medium-sized samples are bolded. This draws the reader’s eye toward the most statistically robust results.

• Sub-national Distinction: We use italics specifically for “Urban Nigeria” and “Rural Nigeria” to indicate that these are sub-divisions of a single national dataset, distinguishing them from the standard country-level points.

To implement this hierarchy, we first map the font styles to our dataset and then register the required Barlow and Caveat font families using relative paths to ensure project portability.

data_manipulated <- data_manipulated %>%
  mutate(text_style = case_when(
      Country %in% c("Urban Nigeria", "Rural Nigeria") ~ "italic", 
      Respondents > 2200 ~ "bold.italic", 
      Respondents >= 1500 & Respondents <= 2200 ~ "bold", 
      TRUE ~ "plain"
    ))

# 3. UNIR el estilo a la tabla de etiquetas ANTES de cambiar nombres
label_positions <- label_positions %>%
  left_join(data_manipulated %>% select(Country, text_style), by = "Country") %>%
  mutate(text_style = replace_na(text_style, "plain")) # Seguridad

# 4. Cargar Fuentes
base_path <- "fonts/"
sysfonts::font_add(family = "Barlow", 
                   regular = paste0(base_path, "Barlow-Regular.ttf"), 
                   bold = paste0(base_path, "Barlow-SemiBold.ttf"),
                   bolditalic = paste0(base_path, "Barlow-BoldItalic.ttf"),
                   italic = paste0(base_path, "Barlow-Italic.ttf"))
sysfonts::font_add(family = "Caveat", 
                   regular = paste0(base_path, "Caveat-Regular.ttf"))
showtext::showtext_auto()

Label aesthetics continue here, now, to ensure the labels are perfectly legible, we apply final adjustments to the text strings and their spatial anchors. This involves splitting long country names into multiple lines using the newline character (\n) and manually setting horizontal (h_align) and vertical (v_align) alignment values. These specific parameters prevent text from overlapping with data points or other labels, ensuring a professional and polished visual output.

# Splitting long country names for better spatial distribution
label_positions <- label_positions %>%
  mutate(
    Country = case_when(
      Country == "São Tomé and Príncipe" ~ "São Tomé\nand Príncipe",
      Country == "Urban Nigeria" ~ "Urban\nNigeria",
      Country == "Rural Nigeria" ~ "Rural\nNigeria",
      Country == "Burkina Faso" ~ "Burkina\nFaso",
      TRUE ~ Country
    )
  )

# Manually assigning alignment anchors (0 = Left/Bottom, 1 = Right/Top)
label_positions$h_align <- 0.5
label_positions$h_align[grepl("Sao Tomé|Príncipe", label_positions$Country)] <- 0
label_positions$h_align[grepl("Burkina|Faso", label_positions$Country)] <- 1

label_positions$v_align <- 0.5
label_positions$v_align[grepl("Sao Tomé|Príncipe", label_positions$Country)] <- 0.3
label_positions$v_align[grepl("Liberia", label_positions$Country)] <- 0.63

Data Segmentation for Layered Visualization

To achieve precise control over the visual elements of our chart, we split the main dataset into two specialized subsets. This segmentation allows us to apply different aesthetic rules to national-level data versus sub-national groups:

• National Dataset: Contains the primary data points for most countries, using standard centered alignments.

• Sub-national Dataset: Specifically isolates “Urban” and “Rural” Nigeria. This allows us to apply unique multiline labeling and specific positioning to these points, highlighting the internal demographic contrast within the same country.

By filtering these groups into separate objects, we can layer them independently within ggplot2, ensuring that labels and points do not overlap and that the visual hierarchy remains intact.

# Creating the primary national dataset
data_countries_m <- data_manipulated |>
  filter(!Country %in% c("Urban Nigeria", "Rural Nigeria")) |>
    mutate(
    h_align = 0.5,
    v_align = 0.5)

# Creating the sub-national dataset for Nigeria  
data_subgroups_m <- data_manipulated |>
  filter(Country %in% c("Urban Nigeria", "Rural Nigeria")) |>
  mutate(
    label_x = P_Value,
    label_y = OR,
    h_align = 0.5,
    v_align = 0.5,
    Country_label = case_when(
      Country == "Urban Nigeria" ~ "Urban\nNigeria",
      Country == "Rural Nigeria" ~ "Rural\nNigeria"))

# Filtering label positions to exclude specific subgroups
label_positions_countries <- label_positions %>%
  filter(!Country %in% c("Urban Nigeria", "Rural Nigeria", "Urban\nNigeria", "Rural\nNigeria"))

Defining Chart Aesthetics and Constants

Before rendering the final visualization, we establish a set of aesthetic constants and structural parameters. These settings ensure visual consistency across the plot:

• Color Scheme: We define specific hex codes for the vertical divider, the shaded background areas, and the highlight zones.

• Coordinate Scales: We set the breaks for the logarithmic x-axis to ensure the P-values are easily interpretable.

• Targeted Highlighting: A specific list of countries is identified to receive a distinct background shading, helping to visually separate results with lower statistical significance from the rest of the dataset.

Moreover, we map our previously defined remittance levels and color palettes to final variables. This simplifies the syntax in the subsequent plotting code, ensuring that the border colors and remittance groups are correctly linked. This organized approach makes the code more modular and easier to maintain.

# Defining hex colors for structural elements
color_vertical <- "#75A3BF"
color_background_shade <- "#DDDEDF"
color_highlight_shade <- "#D7F0F5"

# Setting geometric and scale constants
y_two_thirds <- 2/3
x_breaks <- c(1.000, 0.500, 0.100, 0.010, 0.001)

# List of countries that will receive a shaded background for visual grouping
countries_with_background <- c("Morocco", "Madagascar", "Senegal", "Sudan", "Guinea", "eSwatini", "Kenya", "South Africa")

# Assigning simplified aliases for plotting clarity
remittance_groups <- remittance_groups_levels
border_palette <- border_colors

Axis Gradient Construction

To enhance the visual communication of statistical significance, we generate a dataset specifically for the horizontal axis. Instead of a solid line, we create a logarithmic gradient composed of 100 small segments.

By calculating these segments on a log scale, we can map a smooth color transition that flows from the least significant results (right) to the most significant results (left). This aesthetic choice mirrors high-end editorial data visualizations, guiding the viewer’s eye across the mathematical spectrum of the chart.

x_inicio <- 1.05
x_fin <- 0.001
y_constante <- 1
num_segmentos <- 100 

gradient_line_data <- data.frame(
  index = seq(0, 1, length.out = num_segmentos)
) %>%
  mutate(
    x_start = 10^(log10(x_inicio) + index * (log10(x_fin) - log10(x_inicio))),
    x_end = 10^(log10(x_inicio) + (index + 1/num_segmentos) * (log10(x_fin) - log10(x_inicio)))
  ) %>%
  mutate(x_end = pmin(x_end, x_fin)) %>% 
  mutate(y_val = y_constante) %>%
  filter(index < 1) 

Similar to the horizontal axis, we construct a segmented dataset for the vertical axis. This line represents the threshold of the Odds Ratio. By dividing the line into 100 segments, we can apply a color gradient that transitions from the bottom to the top of the chart.

We calculate a mid_point variable to ensure the color transition is centered around the horizontal baseline, providing a visual anchor that distinguishes between increased and decreased odds of emigration.

y_inicio <- 0.52
y_fin <- 1.95
x_constante <- 1

num_segmentos <- 100

vertical_axis_data <- data.frame(
  index = seq(0, 1, length.out = num_segmentos)
) %>%
  mutate(
    y_start = y_inicio + index * (y_fin - y_inicio),
    y_end = y_inicio + (index + 1/num_segmentos) * (y_fin - y_inicio)
  ) %>%
  mutate(y_end = pmin(y_end, y_fin)) %>%
  mutate(mid_point = abs(index - 0.5)) %>% 
  mutate(x_val = x_constante) %>%
  filter(index < 1)

First Visualization Construction

Finally, we assemble the first part of the visualization. The code is structured in layers to ensure that structural elements like the significance zones and reference lines are placed behind the data points. We use a logarithmic transformation for both axes to represent the statistical nature of the data accurately.

The chart includes several custom features:

• Blue Striped Zone: Highlights the area where results are not statistically significant (p > 0.05).

• Custom Arrows: Gradient-colored arrows represent the direction of statistical significance and the magnitude of the Odds Ratio.

• Segmented Labels: Using geom_shadowtext to ensure country names are readable against the background and other data points.

p <- ggplot(data_countries_m, aes(x = P_Value, y = OR)) +
  
  # 1. Coordinate Scales
  scale_x_continuous(
    breaks = c(1, 0.5, 0.1, 0.05, 0.01, 0.001), 
    labels = c("1.000", "0.500", "0.100", "0.050", "0.010", "0.001\np-value"),
    trans = c("log10", "reverse")
  ) +
  scale_y_continuous(
    breaks = c(0.5, 0.667, 1, 1.5, 2),
    labels = c("\u00BD", "2/3", "1", "1\u00BD", "2"),
    trans = c("log10"),
    sec.axis = sec_axis(~., name = "Odds ratios", breaks = NULL, labels = NULL)
  ) +

  # 2.1. Blue Striped Non-Significant Zone
  geom_rect_pattern(
    aes(xmin = 0.5, xmax = 1.0, ymin = 0.5, ymax = 2.0),
    fill = "white", pattern = 'stripe', pattern_colour = "#D6EFF5",
    pattern_density = 0.00001, pattern_spacing = 0.0065,      
    colour = "transparent", inherit.aes = FALSE
  ) +

  # 2.2. Horizontal Reference Lines (Y=0.5, Y=1.5, Y=2/3)  
  geom_segment(aes(x = 1.0, y = 0.5, xend = 0.001, yend = 0.5), 
               colour = color_background_shade, linewidth = 0.33) +
  geom_segment(aes(x = 1.0, y = 1.5, xend = 0.001, yend = 1.5),
               colour = color_background_shade, linewidth = 0.33) +
  geom_segment(aes(x = 1.0, y = y_two_thirds, xend = 0.001, yend = y_two_thirds),
               colour = color_background_shade, linewidth = 0.33) + 
  geom_segment(aes(x = 1.0, y = 2, xend = 0.047, yend = 2),
               colour = color_background_shade, linewidth = 0.33) + 

  # 2.3. Vertical Cut-off Lines (X=0.5 and X=0.05)  
  geom_segment(aes(x = 0.5, y = 0.49, xend = 0.5, yend = 2.0),
               linetype = "dashed", colour = color_vertical, linewidth = 0.33) +   
  geom_segment(aes(x = 0.05, y = 0.49, xend = 0.05, yend = 2.3),
               linetype = "dashed", colour = color_vertical, linewidth = 0.33) +
  
  # 2.4. Gradient Axis Components
  geom_segment(data = vertical_axis_data,
               aes(x = x_val, y = y_start, xend = x_val, yend = y_end,
                   alpha = mid_point), colour = "#343132", 
               linewidth = 0.5, inherit.aes = FALSE) +
  scale_alpha_continuous(range = c(0.1, 1), guide = "none") +
  
  # Y-Axis Arrows
  geom_segment(aes(x = 1, y = 1.8, xend = 1, yend = 1.95), 
               colour = "#343132", linewidth = 0.5,
               arrow = arrow(angle = 17, length = unit(0.4, "cm"), 
                             ends = "last", type = "closed")) +
  geom_segment(aes(x = 1, y = 0.56, xend = 1, yend = 0.52), 
               colour = "#343132", linewidth = 0.5,
               arrow = arrow(angle = 17, length = unit(0.4, "cm"), 
                             ends = "last", type = "closed")) +
  
  # X-Axis Gradient Arrow
  geom_segment(data = gradient_line_data,
               aes(x = x_start, y = y_val, xend = x_end, yend = y_val,
                   colour = index), linewidth = 0.5,
               arrow = arrow(angle = 17, length = unit(0.4, "cm"),
                             ends = "last", type = "closed"), 
               inherit.aes = FALSE) +

# 3.1. National Data Points
{
    lapply(remittance_groups, function(group) {
      data_grp <- data_countries_m %>% filter(Remittance_Group == group)
      geom_point(data = data_grp, aes(size = Population, fill = Remittance_Group),
                 shape = 21, colour = border_palette[group], 
                 alpha = 0.9, show.legend = FALSE)
    })
  } +
  
# 3.2. Sub-national Data Layer (Nigeria)
  geom_ellipse(data = data_subgroups_m,
               aes(x0 = P_Value, y0 = OR, a = Population / 580000,
                   b = Population / 2750000, angle = 0),
               linetype = "dotted", colour = "#C1C2C4", fill = NA, 
               linewidth = 0.3) +
  
  # 4. Labeling with background shadows for readability
  geom_shadowtext(data = label_positions_countries, 
                  aes(x = label_x, y = label_y, label = Country, 
                      fontface = text_style, hjust = h_align, vjust = v_align),
                  bg.color = "white", bg.r = 0.15, color = "black",
                  lineheight = 0.15, family = "Barlow", size = 13,            
                  show.legend = FALSE) +
  
  geom_shadowtext(data = data_subgroups_m, 
                  aes(x = label_x, y = label_y, label = Country_label, 
                      fontface = text_style, hjust = h_align, vjust = v_align),
                  bg.color = "white", bg.r = 0.15, color = "#C1C2C4",
                  lineheight = 0.15, family = "Barlow", size = 14,            
                  show.legend = FALSE) +

  # 5. Aesthetic Scales
  scale_fill_manual(values = remittance_colors) +
  scale_colour_gradient(low = "#D7D8D9", high = "#6195B5", guide = "none") +
  scale_size_continuous(range = c(2, 20)) +

  # 6. Global Theme
  theme_minimal() +
  theme(panel.grid = element_blank())

Editorial Annotations

To transform the statistical chart into a narrative visualization, we add layers of rich text annotations. These notes provide crucial context, such as explaining the real-world meaning of an Odds Ratio and highlighting specific national cases like Tanzania or Nigeria.

We use the ggtext package’s geom_richtext to allow for HTML styling (like bolding) within the annotations. Finally, we apply a clean, minimal theme, removing standard grid lines to focus the reader’s attention on the data points and the editorial insights.

text_1 <- "An odds ratio of 2 means the odds of considering<br>emigrating are
twice as high for someone who has<br>experienced hunger as for a person of the
same<br>age and gender who has not experienced hunger."

text_2 <- "The effect of hunger on<br>considerations of emigrating is<br>almost
identical in Côte d'Ivoire<br>and Mali, and the statistical<br>certainty is very
similar. But in<br>the past, the effect in Côte d'Ivoire<br>would have been
dismissed as<br>'not statistically significant'."

text_3 <- "The statistical certainty of the results is<br>much higher in Togo
than in Mali and Côte<br>d'Ivoire, although the size of the effect is<br>similar.
This is mainly because the number<br>of people who consider emigration is
much<br>smaller in the other two countries."

text_4 <- "Lesotho and The Gambia are small<br>countries where emigration is
very<br>widespread. People who experience<br>hunger are much more likely
that<br>others to think about leaving."

text_5 <- "Greater <b>statistical certainty</b><br>that hunger makes a
difference<br>to considerations of emigrating"
  
text_6 <- "Nigeria is a huge and diverse country. Splitting it<br>into the urban
and rural population shows that<br>hunger makes city dwellers much less likely
to<br>consider emigrating, while it hardly makes a<br>difference in rural areas."
  
text_7 <- "Emigration is remarkably rare in Tanzania,<br>whch might explain why
it is very unlikely<br>to be on the mind of people who experience<br>hunger.
Moreover, the survey sample was<br>large, which further boosts<br>statistical
certainty." 

text_8 <- "Conventional<br>(but criticized<br>and abandoned)<br>cut-off for
what<br>is 'statistically<br>significant'"

text_9 <- "Just as likely to<br>be a coincidence<br>as to reflect an<br>actual
relationship<br>between hunger<br>and emigration"

text_10 <- "In many countries,<br>hunger appers<br>to make no<br>difference
to<br>considerations of<br>emigrating. The<br>slight effects we see<br>might be
purely<br>coincidental."

Graph with surronding text

To move beyond a standard scatter plot, we utilize the ggtext package. The function geom_richtext() is essential here because it allows us to render HTML and Markdown. This is what enables the use of tags like <br> for line breaks and <b> for bolding specific words within the annotations, creating a sophisticated typographic flow.

Additionally, we use coord_cartesian(clip = "off"). By default, ggplot2 cuts off any element (like labels or lines) that extends outside the plot area. By turning “clipping” off, we allow our annotations and axis arrows to breathe and occupy the margins, which is a common technique in high-end data journalism to maximize the use of white space.

p_main <- p +
  
# Title and subtitle
  labs(
    title = "Does hunger make Africans want to move abroad?",
    subtitle = "Extensive survey data shows that there is not a simple answer.") +
  
  # Texts
  geom_richtext(
    aes(x = 1.5, y = 1.7, 
        label = "Hunger<br>makes people<br><b>more likely</b><br>to
        consider<br>emigrating"),
    hjust = 1, vjust = 0.5,
    size = 24, lineheight = 0.2,
    fill = NA, color = "black", label.color = NA,
    family = "Barlow") +
  
  geom_richtext(
    aes(x = 1.5, y = 0.6, 
        label = "Hunger<br>makes people<br><b>less likely</b><br>to
        consider<br>emigrating"),
    hjust = 1, vjust = 0.5,
    size = 24, lineheight = 0.2,
    fill = NA, color = "black", label.color = NA,
    family = "Barlow") +
  
    geom_richtext(
    aes(x = 4.5, y = 1.15), 
    label = text_10,
    hjust = 0,            
    vjust = 1,             
    size = 18,            
    lineheight = 0.05,
    fill = NA, 
    label.color = NA,
    color = "gray70",      
    family = "Caveat") +
  
  geom_richtext(
    aes(x = 1, y = 2.25), 
    label = text_1,
    hjust = 0,            
    vjust = 1,             
    size = 18,            
    lineheight = 0.05,
    fill = NA, 
    label.color = NA,
    color = "gray70",      
    family = "Caveat") +
  
  geom_richtext(
    aes(x = 0.0465, y = 2.3),
    label = text_2,
    hjust = 0,             
    vjust = 1,            
    size = 18,            
    lineheight = 0.05,
    fill = NA, 
    label.color = NA,
    color = "gray70",     
    family = "Caveat") +
  
  geom_richtext(
    aes(x = 0.0045, y = 2.28),
    label = text_3,
    hjust = 0,            
    vjust = 1,            
    size = 18,          
    lineheight = 0.05,
    fill = NA, 
    label.color = NA,
    color = "gray70", 
    family = "Caveat") +
  
  geom_richtext(
    aes(x = 0.004, y = 1.48),
    label = text_4,
    hjust = 0,            
    vjust = 1,         
    size = 18,          
    lineheight = 0.05,
    fill = NA, 
    label.color = NA,
    color = "gray70",     
    family = "Caveat") +
  
  geom_richtext(
    aes(x = 0.00132, y = 0.853),
    label = text_5,
    hjust = 1,
    vjust = 0,
    size = 24,
    lineheight = 0.17,
    fill = NA, 
    label.color = NA,
    color = "#6195B5",
    family = "Barlow") +
  
  geom_richtext(
    aes(x = 0.0094, y = 0.66),
    label = text_6,
    hjust = 0,
    vjust = 0,
    size = 18,
    lineheight = 0.05,
    fill = NA, 
    label.color = NA,
    color = "gray70",
    family = "Caveat") +
  
  geom_richtext(
    aes(x = 0.007, y = 0.55),
    label = text_7,
    hjust = 0,
    vjust = 0,
    size = 18,
    lineheight = 0.05,
    fill = NA, 
    label.color = NA,
    color = "gray70",
    family = "Caveat") +
  
  geom_richtext(
    aes(x = 0.048, y = 0.51),
    label = text_8,
    hjust = 0,
    vjust = 0,
    size = 19,
    lineheight = 0.18,
    fill = NA, 
    label.color = NA,
    color = "#6195B5",
    family = "Barlow") +
  
    geom_richtext(
    aes(x = 0.48, y = 0.51),
    label = text_9,
    hjust = 0,
    vjust = 0,
    size = 19,
    lineheight = 0.18,
    fill = NA, 
    label.color = NA,
    color = "#6195B5",
    family = "Barlow") +

  coord_cartesian(clip = "off") +
  theme(
    plot.margin = margin(t = 5, r = 20, b = 5, l = 40, unit = "pt"),
    plot.title = element_text(size = 120, face = "bold", family = "Barlow"),
    plot.subtitle = element_text(size = 80, family = "Barlow"),
    axis.line = element_blank(),
    axis.ticks = element_blank(),
    panel.grid = element_blank(),
    axis.text.x = element_text(size = 35, margin = margin(t = -20, unit = "pt")),
    axis.text.y = element_text(size = 35, margin = margin(r = -20, unit = "pt")),
  guides(
    fill = "none",  
    size = "none", 
    colour = "none"))

Creating the legends

While our graphic already displays the core findings, the bottom panel of our infographic serves as the “decoding key.” Instead of relying on standard, automatically generated legends, we build custom visual components to ensure they match the original illustration.

This legend system is divided into four strategic modules:

1. Remittance Intensity: A vertical scale that maps colors to the share of people that declare receiving remittances, providing a cross-national economic context.

2. Sample Size and Statistical Confidence: A guide explaining how font styles (bold, italic) relate to the number of respondents.

3. Population Scaling: A custom semi-circle legend that allows the reader to estimate country sizes at a glance.

4. Academic Sourcing and Notes: A dedicated space for transparency, detailing the data origins, survey questions, and credits.

To achieve this layout, we use the patchwork package to combine these independent plots with the main chart, ensuring that all typographic and color elements remain perfectly synchronized across the entire document.

Remittances

In this chunk, we manually construct the legend using a dedicated dataframe. Key steps include:

• Coordinate Mapping: We define precise y positions for each categorical level to ensure equal vertical spacing.

• Identity Scales: By using scale_fill_identity() and scale_color_identity(), we tell R to use the hex codes stored directly in our dataframe, maintaining a perfect match with the country bubbles in the main chart.

• Visual Separators: We add small horizontal segments between categories to enhance scannability.

• Contextual Commentary: We use the handwritten-style Caveat font to add a brief sociological explanation of why remittances matter in the context of food insecurity and migration.

library(dplyr)
library(stringr)

# 1. DATAFRAME FOR THE LEGEND
remittance_data <- data.frame(
  percent = remittance_groups_levels,
  y = length(remittance_groups_levels):1)

remittance_data <- remittance_data %>%
  mutate(
    fill_color = remittance_colors[percent], 
    border_color = border_colors[percent]) %>%
  mutate(
    fill_color = as.character(fill_color),
    border_color = as.character(border_color))

remittance_groups_levels <- c("> 40%", "35-40%", "30-35%", "25-30%", "20-25%", 
                               "15-20%", "10-15%", "< 10%")

## SETTING AESTHETICS OF THE LEGEND
y_step <- 0.31
num_items <- length(remittance_groups_levels)

remittance_data <- data.frame(
  percent_raw = remittance_groups_levels,
  y = seq(from = num_items * y_step, 
          to = y_step, 
          by = -y_step)) %>%
  mutate(
  percent_label = case_when(
    percent_raw == "> 40%" ~ "40%",
    percent_raw == "35-40%" ~ "35%",
    percent_raw == "30-35%" ~ "30%",
    percent_raw == "25-30%" ~ "25%",
    percent_raw == "20-25%" ~ "20%",
    percent_raw == "15-20%" ~ "15%",
    percent_raw == "10-15%" ~ "10%",
    TRUE ~ NA_character_)) %>%
  mutate(
    fill_color = remittance_colors[percent_raw], 
    border_color = border_colors[percent_raw]) %>%
  mutate(
    fill_color = as.character(fill_color),
    border_color = as.character(border_color))

line_data <- data.frame(
  y = seq(from = (num_items - 1) * y_step + y_step / 2, 
          to = y_step * 1.5,                             
          by = -y_step))

explanatory_text <- "If more people receive\nmoney transfers from\nrelatives or
friends\nabroad, it probably\nmeans that emigration\nis more widespread—\nand
thereby more\nrelatable as a possible\nway of hardship. If\nemigration is rare,
it\nmight be more of\nanelite phenomenon,\nbeyond the imagination\nof those who
experience\nhunger."

# CREATION GRAPH

p_legend_remittance <- ggplot() +
  
  # Separator segments
  geom_segment(data = line_data, aes(x = 0.193, y = y + 0.6, xend = 0.35,
                                     yend = y + 0.6), linewidth = 0.2,
               color = "black", inherit.aes = FALSE) +
  
  # Legend points
  geom_point(data = remittance_data, aes(x = 0.3, y = y+0.6, fill = fill_color,
                                         color = border_color), shape = 21,
             size = 3.5, linewidth = 0.5) + 
  
  geom_text(data = remittance_data, aes(x = 0.05, y = y + 0.46,
                                        label = percent_label), hjust = 0,
            size = 11, family = "Barlow") +
  
  # Title
  geom_text(aes(x = 0.05, y = 3.8, label = "Share of people\nwho
                receive\nremittances"), 
            vjust= 0.7, hjust = 0, lineheight = 0.18, fontface = "bold",
            size = 11, family = "Barlow") +
  
  geom_segment(aes(x = 0.05, y = 3.35, xend = 1, yend = 3.35), vjust = 0,
               linewidth = 0.4, color = "black") + 
  
  # Descriptive sidebar
  geom_text(aes(x = 0.43, y = 1.8, label = explanatory_text), 
            hjust = 0, vjust = 0.6, lineheight = 0.13, fontface= "plain",
            size = 16, color = "gray50", family = "Caveat") +

  # Structure
  scale_fill_identity() + 
  scale_color_identity() + 
  
  coord_cartesian(
    xlim = c(0, 1.2), 
    ylim = c(0, 4.2), 
    clip = "off") +
  
theme_void() + 
  theme(
    plot.margin = margin(0.0, 0, 0, 0, unit = "cm"),
    plot.title = element_text(margin = margin(0, 0, 0, 0, unit = "pt")))

Country Label

The next module explains the relationship between sample size and typographic encoding. In this section, we provide a key for the reader to understand that the different font weights used for the country names in the scatter plot are not merely decorative but represent the statistical robustness of the data.

In this chunk, we define a legend for the number of respondents. We use three distinct fontface settings:

• Bold Italic: Highest sample sizes (2200-2400), representing the greatest statistical certainty.

• Bold: Mid-range sample sizes (1500-1800).

• Plain: Standard sample sizes (1100-1200).

By layering multiple geom_text calls at specific y coordinates, we create a structured list that functions as a manual legend. An accompanying sidebar provides a comparative example (Uganda vs. Ghana) to explain why a larger sample size improves confidence in the results.

# 1. Defining explanatory text
text_label <- "Larger samples yield\ngreater statistical\ncertainty. So,
although\nthe effect of hunger is\nstronger in Uganda\nthan in Ghana, we can\nbe
more confident that\nthere really is an\neffect in Ghana, simply\nbecause the
survey in\nGhana had twice as\nmany respondents."

# 2. Building the respondents legend plot
p_respondents <- ggplot() +
geom_text(aes(x = 0.05, y = 4.05, label = "Number of\nrespondents\nin the
              survey\n(sample size)"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "bold", size = 11,
          family = "Barlow") +
  
  # Separator Line
geom_segment(aes(x = 0.05, y = 3.4, xend = 1, yend = 3.4), 
             linewidth = 0.4, color = "black") +

  # Typographic samples (Font face mapping)
  geom_text(aes(x = 0.05, y = 3.2, label = "Country"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "bold.italic",
          size = 11, family = "Barlow") +
    geom_text(aes(x = 0.05, y = 3.02, label = "Country"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "bold",
          size = 11, family = "Barlow") +
      geom_text(aes(x = 0.05, y = 2.82, label = "Country"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "plain",
          size = 11, family = "Barlow") +
  
  # Sample size numeric labels
      geom_text(aes(x = 0.33, y = 3.2, label = "2200-2400"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "plain",
          size = 11, family = "Barlow") +
  
      geom_text(aes(x = 0.33, y = 3.02, label = "1500-1800"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "plain",
          size = 11, family = "Barlow") +
  
       geom_text(aes(x = 0.33, y = 2.82, label = "1100-1200"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "plain",
          size = 11, family = "Barlow") +
  
  # Side explanatory annotation 
  geom_text(aes(x = 0.05, y = 3.3, label = text_label), 
            hjust = 0, vjust = 1.4, lineheight = 0.13, fontface= "plain",
            size = 16, color = "gray50", family = "Caveat") +
  
scale_fill_identity() + 
scale_color_identity() + 
coord_cartesian(
  xlim = c(0, 1.2), 
  ylim = c(0, 4.2),
  clip = "off") +
theme_void() + 
theme(
  plot.margin = margin(0.0, 0, 0, 0, unit = "cm"),
  plot.title = element_text(margin = margin(0, 0, 0, 0, unit = "pt")))

Country Size

The third module of the footer panel addresses Population Scaling. In the main scatter plot, the size of each bubble corresponds to the country’s population. To help the reader calibrate this scale, we implement a custom-built semi-circle legend.

Unlike standard legends, this component is created by calculating trigonometric paths to draw concentric semi-circles. This design is space-efficient and follows high-end data journalism aesthetics.

• Trigonometric Mapping: We use a sequence of angles (theta) to generate the coordinates for each semi-circle, which are then rendered using geom_polygon.

• Scale Synchronization: The scale_factor is carefully calibrated to match the range used in the main plot’s scale_size_continuous, ensuring the visual representation is accurate.

• Neutral Palette: We use a monochromatic gray scale for the legend to avoid visual competition with the remittance color coding.

# 1. Defining the scaling parameters and explanatory notes
scale_factor <- 0.1385
explanatory_text_size <- "Population size has no\nimpact on
statistical\nuncertainty, but it\nmight be that people in\nsmall countries
are\nmore inclined to see\nemigration as a\npathway out of hunger."

x_lim <- c(-1.5, 1.2)
y_lim <- c(0, 4.2)

# 2. Generating the semi-circle geometry
radios <- c(4.3, 3, 1.5, 0.6)
offset_x <- -0.85
offset_y <- 2.24

# 3. setting colors_semi_circles
colors_semi_circles <- c(
  "0.6"   = "#A9A9AB",
  "1.5"  = "#B7B8B9",
  "3"  = "#CACBCC",
  "4.3" = "#E2E2E3"
)

# 4. Semicircles
datos <- do.call(rbind, lapply(radios, function(r) {
  theta <- seq(-pi/2, pi/2, length.out = 300)
  data.frame(
    x = offset_x - (r * scale_factor) * cos(theta),
    y = offset_y + (r * scale_factor) * sin(theta),
    r = factor(r, levels = radios)
  )
}))
  
# 5. Plotting the custom legend
p_legend_size <- ggplot(datos, aes(x, y, group = r, fill = r)) +
  geom_polygon(color = "#8F8D8E", linewidth = 0.2) +
  
    geom_text(
    aes(x = -1.5, y = 3.5, label = "Population"),
    hjust = 0, fontface = "bold", family = "Barlow", size = 13
  ) +
  
  geom_segment(aes(x = -1.5, y = 3.34, xend = 0.5, yend = 3.34), 
             linewidth = 0.4, color = "black") +
  
  geom_segment(aes(x = -0.85, y = 2.798, xend = -0.47, yend = 2.798), 
             linewidth = 0.22, color = "#8F8D8E") +
  
  geom_text(aes(x = -0.45, y = 2.88, label = "200 million"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "plain",
          size = 11, family = "Barlow") +
  
  geom_segment(aes(x = -0.85, y = 2.621, xend = -0.47, yend = 2.621), 
             linewidth = 0.22, color = "#8F8D8E") +
  
  geom_text(aes(x = -0.445, y = 2.636, label = "80 million"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "plain",
          size = 11, family = "Barlow") +
  
  geom_segment(aes(x = -0.85, y = 2.412, xend = -0.47, yend = 2.412), 
             linewidth = 0.22, color = "#8F8D8E") +
  
  geom_text(aes(x = -0.445, y = 2.46, label = "10 million"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "plain",
          size = 11, family = "Barlow") +
  
  geom_segment(aes(x = -0.85, y = 2.206, xend = -0.47, yend = 2.206), 
             linewidth = 0.22, color = "#8F8D8E") +
  
  geom_text(aes(x = -0.445, y = 2.24, label = "1 million"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "plain",
          size = 11, family = "Barlow") +
  
  geom_text(
    aes(x = -1.5, y = 1.2, label = explanatory_text_size),
    hjust = 0, vjust = 1, lineheight = 0.15, 
    fontface = "italic", size = 11, color = "gray50", family = "Caveat"
  ) +
  
  scale_fill_manual(values = colors_semi_circles) +
  coord_equal(
    xlim = x_lim,
    ylim = y_lim,
    expand = FALSE
  ) +
  theme_void() +
  theme(legend.position = "none")

Sources

The final component of our infographic footer is the Sources and Notes section. Transparency in data journalism is vital, especially when dealing with statistical models like logit regressions and sensitive survey data regarding food insecurity and migration.

This module provides the necessary academic and technical context for the entire visualization. It details the Afrobarometer Round 7 origin, the specific survey question IDs (Q68A and Q8D), and the regression controls used (age and gender).

By creating this as a separate ggplot object, we can treat the metadata as a design element, ensuring that long strings of text are wrapped and positioned to balance the visual weight of the other three legend modules.

# 1. Defining the technical and academic metadata
text_sources <- "See jorgencarling.org/2024/12/31/hunger-and-migration for a full
explanation.\n\nData source: Afrobarometer Round 7. Data collected in 2019 (the
most recent year for\nwhich data on considerations of emigration are available.)
Question formulations:\nQ68A 'How much,if at all, have you considered to moving
to another country to live?'\n('A lot': 1; Other non-missing answers: 0).
Q8A'Over the past year, how often, if ever,\nhave you or anyone in your family
gone without enough food to eat? ('Never': 0; Other\nnon-missing answers:
1).\n\nRegression analysis: logit regression using Stata's presets for survey
data analysis.\nAge and gender are included as controls.\n\nCredits: Data
analysis and visualization by Jørgen Carling, 2024. License: CC-BY.\nCreated in
conjunction with the project Future Migration as Present Fact (FUMI),\nfunded by
the European Research Council, grant agreement n° 819227, and carried\nout at
the PRIO Migration Centre. See prio.org/fumi."

# 2. Building the sources plot
p_sources <- ggplot() +

  # Section Title
  geom_text(aes(x = 0, y = 3.6, label = "Sources and notes"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "bold", size = 11, family = "Barlow") +
  
  # Structural Separator
  geom_segment(aes(x = 0, y = 3.4, xend = 1.15, yend = 3.4), 
             linewidth = 0.4, color = "black") + 
  
  # Metadata Body
  geom_text(aes(x = 0, y = 3.1, label = text_sources),
          vjust = 1, 
          hjust = 0, lineheight = 0.2, fontface = "plain", size = 9, family = "Barlow") +
  
scale_fill_identity() + 
scale_color_identity() + 
coord_cartesian(
  xlim = c(0, 1.2), 
  ylim = c(0, 4.2),
  clip = "off") +
theme_void() + 
theme(
  plot.margin = margin(0.0, 0, 0, 0, unit = "cm"),
  plot.title = element_text(margin = margin(0, 0, 0, 0, unit = "pt")))

Combination and final replication graph

To conclude the replciation, we assemble the individual components into a single, high-resolution infographic. This is achieved using the patchwork package, which provides an intuitive syntax for combining multiple ggplot objects into complex layouts.

The assembly follows a hierarchical structure:

• The Bottom Row: We first group the four legend modules (p_respondents, p_legend_size, p_legend_remittance, and p_sources) using the | operator. We use plot_layout(widths = ...) to manually define the horizontal space each module occupies, ensuring the source notes have enough width for readability.

• The Vertical Stack: Using the / operator, we place the main scatter plot (p_combined) on top of the newly created bottom row. We set a heights ratio of 3:1 to give the data visualization the prominence it deserves while keeping the legends accessible.

• Unified Styling: The & operator is used to apply theme adjustments (like margins) across all sub-plots simultaneously, ensuring a perfectly aligned and professional border for the final output.

# 1. Adjusting margins for the main chart to align with the footer
p_combined <- p_main + theme(
    plot.margin = margin(t = 5, r = 5, b = 5, l = 20) # Aumentar 'l' (izquierda)
)

# 2. Assembling the footer row
bottom_row_final <- (p_respondents | p_legend_size | p_legend_remittance | p_sources) +
  plot_layout(widths = c(
    1.73,
    1.73,
    1.73,  
    3.51     
  )) +
  theme(
    panel.spacing = unit(0, "pt"))

# 3. Final vertical composition (Main Chart / Footer)
final_plot <- p_combined / bottom_row_final

# 4. Final layout scaling and annotation theme
final_plot <- final_plot +
  plot_layout(heights = c(3, 1)) +
  plot_annotation(theme = theme(plot.margin = margin(t = 5, r = 5, b = 5, l = 5)) & 
  theme(plot.margin = margin(t = 5, r = 5, b = 5, l = 5)))

Replication

Enhancing Visual Clarity in Complex Survey Data

The original visualization provided a foundational look at the relationship between hunger and migration. However, to transform it into a publication-ready “Forest Plot,” several strategic design improvements were implemented to reduce cognitive load and enhance interpretability.

1. Reducing Visual Noise and Redundancy

• Removal of Population Weighting: The previous version included a variable related to population size that added unnecessary complexity. By removing this dimension, the focus shifts entirely to the Odds Ratios (OR) and their statistical significance, which are the primary drivers of the research question.

• Cleaner Background: Using a minimal theme with light gridlines ensures that the data points (the markers) are the focal point, not the coordinate system.

2. Implementation of a Standardized Forest Plot

• Logarithmic Scaling: The x-axis now utilizes a log scale. This is statistically superior for Odds Ratios because it treats “less likely” and “more likely” effects symmetrically around the 1.0 null line.

• Country Identification: By reordering countries based on their coefficients and providing a clear vertical layout, the reader can instantly recognize which nations show the strongest correlations without visual scanning fatigue.

3. Precision and Uncertainty (Error Bars)

• 95% Confidence Intervals: Every point is accompanied by explicit error bars.

• Visual Encoding of Significance: A dual-tone system was introduced. Results that are statistically significant are highlighted in a darker, “high-ink” shade, while non-significant results (those crossing the 1.0 line) are rendered in a lighter shade. This allows the viewer to filter out “statistical noise” at a glance.

4. Contextual Nuance: The Case of Nigeria

• Disaggregation: Nigeria is presented separately for urban and rural areas. Given its massive demographic weight and the sharp internal differences in migration drivers, this disaggregation prevents a misleading “national average” and provides a more honest representation of the data.

5. Integrated Annotations and Legends

• Self-Explanatory Design: Instead of relying on a separate caption, the legend is integrated into the layout. This includes a technical explanation of the Odds Ratios and the “Share of people receiving remittances” (represented by color intensity), making the graphic a standalone piece of scientific communication.

Charging the required data

As in the replication, we have all results prepared in this other R document, that contains country with its results for the coefficients as their confidence intervals.

source("alternative_results.R")

Adjustments for the graph

In this stage, we focus on data restructuring to optimize the visual hierarchy of the forest plot:

• Logical Reordering: Countries are no longer listed alphabetically. Instead, they are reordered by their Odds Ratio (OR) values. This allows the reader to immediately distinguish the gradient of hunger’s impact across the continent, from the strongest to the weakest correlations.

• Aesthetic Categorization: We transformed the continuous remittance data into discrete groups. This categorization is essential for the color-coded legend, helping to identify patterns between external financial support (remittances) and migration desires more clearly.

• Semantic Labeling: We defined rich-text labels (text_more_likely and text_less_likely) to provide clear, intuitive anchors at the ends of the X-axis. This guides the reader’s interpretation of the statistical coefficients without requiring deep prior knowledge of log-scale Odds Ratios.

# 1. Data preparation and remittance group assignment
alt_data_plot <- alt_data_graph %>%
  mutate(
    Remittance_Group = case_when(
      alt_Remittances > 40 ~ "> 40%",
      alt_Remittances > 35 & alt_Remittances <= 40 ~ "35-40%",
      alt_Remittances > 30 & alt_Remittances <= 35 ~ "30-35%",
      alt_Remittances > 25 & alt_Remittances <= 30 ~ "25-30%",
      alt_Remittances > 20 & alt_Remittances <= 25 ~ "20-25%",
      alt_Remittances > 15 & alt_Remittances <= 20 ~ "15-20%",
      alt_Remittances > 10 & alt_Remittances <= 15 ~ "10-15%",
      alt_Remittances <= 10 ~ "< 10%",
      TRUE ~ "NA"
    ),
    # Ensure factor levels follow a logical descending order for the legend
    Remittance_Group = factor(Remittance_Group,
                              levels = remittance_groups_levels),
    # Reorder countries based on the Odds Ratio (OR) value to create the
    # forest plot structure
    Country = reorder(Country, alt_OR)
  )

# Define descriptive HTML-formatted labels for the plot's X-axis extremes
text_more_likely <- "Hunger<br>makes people<br><b>more likely</b><br>to
consider<br>emigrating"
text_less_likely <- "Hunger<br>makes people<br><b>less likely</b><br>to
consider<br>emigrating"

Creation of the main visualization

This chunk represents the core of the visual redesign, where the “Forest Plot” is meticulously constructed to prioritize statistical clarity:

• Log-Scale Symmetry: By using scale_x_log10, we ensure that an Odds Ratio of 0.5 (half as likely) is visually equidistant from the null line (1.0) as an OR of 2.0 (twice as likely). This prevents the visual bias common in linear scales where large values appear more significant than small ones.

• Significance Highlighting: We implemented a “de-emphasis” strategy. Results that cross the 1.0 threshold (not statistically significant) are rendered with higher transparency and lighter colors. This allows the significant “real world” effects to pop out, reducing the time required to interpret the findings.

• Nigeria Disaggregation: Urban and Rural Nigeria are plotted with neutral gray tones. This stylistic choice acknowledges their unique status as sub-national data points while preventing their specific data from overwhelming the color-coded “National” results.

• Minimalist Tufte-style Theme: We removed unnecessary borders and heavy gridlines. The use of the “Barlow” typeface and clean gray95 lines follows the principle of maximizing the data-to-ink ratio, ensuring every element on the screen serves a purpose.

alt_p_forest <- ggplot(alt_data_plot, aes(x = alt_OR, y = Country)) +
  # Baseline (OR = 1): Represents the null hypothesis (no effect)
  geom_vline(xintercept = 1, linetype = "longdash", color = "gray70",
             linewidth = 0.2) +
  
  # Error Bars: Differentiate significance using transparency (alpha) and color
  geom_errorbarh(aes(xmin = alt_CI_Lower, xmax = alt_CI_Upper, 
                     alpha = alt_is_significant,
                     color = alt_is_significant),
                 height = 0.4, linewidth = 0.3) +
  
  # Manual scales for error bars
  scale_color_manual(values = c("Statistically Significant" = "gray20", 
                                "Not Significant" = "gray85"), 
                     guide = "none") +
  
  scale_alpha_manual(values = c("Statistically Significant" = 1, 
                                "Not Significant" = 0.5), 
                     guide = "none") +
  
  geom_point(data = filter(alt_data_plot, !Country %in% c("Urban Nigeria",
                                                          "Rural Nigeria")),
             aes(fill = Remittance_Group, color = Remittance_Group),
             shape = 21, size = 3, stroke = 0.35) +
  
  geom_point(data = filter(alt_data_plot, Country %in% c("Urban Nigeria",
                                                         "Rural Nigeria")),
             fill = "#DDDEDF", color = "#8F8D8E", 
             shape = 21, size = 3, stroke = 0.4, alpha = 0.6) +

  # Axis and Scale configuration
  scale_fill_manual(values = remittance_colors) +
  scale_color_manual(values = border_colors) +
  scale_x_log10(limits = c(0.34, 3.03), breaks = c(0.5, 1, 2),
                labels = c("0.5", "1.0", "2.0")) +
  
  # Titles and removing automatic legends
  labs(title = "Does hunger make African want to move abroad?",
       subtitle = "Extensive survey data shows that there is not a
       simple answer.",
       x = "Odds Ratio (Log Scale)", y = "") +
  
  scale_y_discrete(expand = expansion(add = c(0.8, 0.2))) +
  
  
  # Theme Configuration
  theme_minimal() +
  theme(
    # Text Alignment: Left-align title and subtitle
  plot.title = element_text(size = 35, face = "bold", hjust = 0),
  plot.subtitle = element_text(
    size = 27, 
    face = "italic", 
   hjust = 0, 
   margin = margin(b = 13)
  ),
  
  # Alignment with the plot edge (not just the panel)
  plot.title.position = "plot",
    text = element_text(family = "Barlow"),
    axis.text.y = element_text(size = 14, face = "plain", color = "gray40"),
    axis.text.x = element_text(size = 14, family = "Barlow", color = "gray40"),
    panel.grid.major.y = element_line(linewidth = 0.17, color = "gray95"), 
    panel.grid.major.x = element_line(linewidth = 0.17, color = "gray95"), 
    panel.grid.minor = element_blank(), 
    axis.line.x = element_line(color = "gray95", linewidth = 0.17),
    axis.ticks.x = element_line(color = "gray95", linewidth = 0.17),
    axis.ticks.length.x = unit(3, "pt"),
    axis.text.x.top = element_text(margin = margin(t = 1)),
    legend.position = "none",
    plot.margin = margin(t = 12, r = 12, b = 12, l = 12)
  ) +
  coord_cartesian(clip = "off")

Adding the additional text

Here, we add the additional explanatory texts to enhance comprehension, but not as much as text the original graph in order to avoid visual noise.

alt_main <- alt_p_forest +
  scale_y_discrete(expand = expansion(add = c(0.6, 0.2))) +
  annotate(
    geom = "richtext",        
    x = 0.46, 
    y = 0.4,                 
    label = text_less_likely, 
    family = "Barlow", 
    size = 5,
    lineheight = 0.4,
    color = "black", 
    fill = NA,                
    label.color = NA,         
    vjust = 1,
    hjust = 1
  ) +
    annotate(
    geom = "richtext",        
    x = 2.16, 
    y = 0.4,                 
    label = text_more_likely, 
    family = "Barlow", 
    size = 5,
    lineheight = 0.4,
    color = "black", 
    fill = NA,                
    label.color = NA,         
    vjust = 1,
    hjust = 0
  ) +
  theme (
    plot.margin = margin(t = 12, r = 12, b = 20, l = 12)
  )

Legend creation

Since standard automatic legends often lack the necessary detail for complex social science data, we built custom, modular legends to provide deep context and transparency.

Remittances

We will use the same legend setting for remittances as colors are conserved, it is only needed adjusting for size.

# 1. DATAFRAME FOR THE LEGEND
remittance_data <- data.frame(
  percent = remittance_groups_levels,
  y = length(remittance_groups_levels):1)

remittance_data <- remittance_data %>%
  mutate(
    fill_color = remittance_colors[percent], 
    border_color = border_colors[percent]) %>%
  mutate(
    fill_color = as.character(fill_color),
    border_color = as.character(border_color))

remittance_groups_levels <- c("> 40%", "35-40%", "30-35%", "25-30%", "20-25%", 
                               "15-20%", "10-15%", "< 10%")

## SETTING AESTHETICS OF THE LEGEND
y_step <- 0.32
num_items <- length(remittance_groups_levels)

remittance_data <- data.frame(
  percent_raw = remittance_groups_levels,
  y = seq(from = num_items * y_step, 
          to = y_step, 
          by = -y_step)) %>%
  mutate(
  percent_label = case_when(
    percent_raw == "> 40%" ~ "40 %",
    percent_raw == "35-40%" ~ "35 %",
    percent_raw == "30-35%" ~ "30 %",
    percent_raw == "25-30%" ~ "25 %",
    percent_raw == "20-25%" ~ "20 %",
    percent_raw == "15-20%" ~ "15 %",
    percent_raw == "10-15%" ~ "10 %",
    TRUE ~ NA_character_)) %>%
  mutate(
    fill_color = remittance_colors[percent_raw], 
    border_color = border_colors[percent_raw]) %>%
  mutate(
    fill_color = as.character(fill_color),
    border_color = as.character(border_color))

line_data <- data.frame(
  y = seq(from = (num_items - 1) * y_step + y_step / 2, 
          to = y_step * 1.5,                             
          by = -y_step))

explanatory_text <- "If more people receive\nmoney transfers from\nrelatives or
friends\nabroad, it probably\nmeans that emigration\nis more widespread—\nand
thereby more\nrelatable as a possible\nway of hardship. If\nemigration is rare,
it\nmight be more of\nanelite phenomenon,\nbeyond the imagination\nof those who
experience\nhunger."


# CREATION GRAPH

p_legend_remittance_alt <- ggplot() +
  geom_segment(data = line_data, aes(x = 0.23, y = y + 0.6, xend = 0.388,
                                     yend = y + 0.6),
               linewidth = 0.16, color = "black", inherit.aes = FALSE) +
  geom_point(data = remittance_data, aes(x = 0.35, y = y+0.6, fill = fill_color,
                                         color = border_color),
             shape = 21, size = 2.35, linewidth = 0.05) + 
  
  geom_text(data = remittance_data, aes(x = 0.02, y = y + 0.44,
                                        label = percent_label),
            hjust = 0, size = 5.5, family = "Barlow") +
  
  geom_text(aes(x = 0, y = 4, label = "Share of people\nwho receive\nremittances"), 
            vjust= 0.7, hjust = 0, lineheight = 0.3, fontface = "bold",
            size = 6, family = "Barlow") +
  
  geom_segment(aes(x = 0, y = 3.41, xend = 0.9, yend = 3.41), vjust = 0,
               linewidth = 0.3, color = "black") + 
   
  geom_text(aes(x = 0.52, y = 1.75, label = explanatory_text), 
            hjust = 0, vjust = 0.6, lineheight = 0.24, fontface= "plain",
            size = 6.3, color = "gray70", family = "Caveat") +

  scale_fill_identity() + 
  scale_color_identity() + 
  
  coord_cartesian(
    xlim = c(0, 1.2), 
    ylim = c(0, 4.2), 
    clip = "off") +
  
theme_void() + 
  theme(
    plot.background = element_rect(fill = "white", color = NA),
  panel.background = element_rect(fill = "white", color = NA),
    plot.margin = margin(0.0, 0, 0, 0, unit = "cm"),
    plot.title = element_text(margin = margin(0, 0, 0, 0, unit = "pt")))

Sources

This component ensures full transparency regarding data origins and methodological constraints.

text_sources_alt <- "<b>Full explanation:</b>
jorgencarling.org/2024/12/31/hunger-and-migration<br><br><b>Data source:</b>
Afrobarometer Round 7 (2019). Survey data covering<br> considerations of
emigration (Q68A) and food insecurity (Q8D).<br><br><b>Methodology:</b> Results
based on logit regression analysis, controlling<br> for age and gender. Results
are shown a log scale for Odds Ratios.<br><br><b>Technical note:</b> Error bars
represent 95% confidence intervals.<br><br><b>Analysis:</b> Nigeria is
disaggregated by urban and rural areas due to its<br>large population weight and
significant demographic differences.<br><br><b>Credits:</b> Original research
and analysis by Jørgen Carling (2024).<br> Visualization improved by María
Garcés Blázquez. Part of the FUMI<br> project at the PRIO Migration Centre."

p_sources_alt <- ggplot() +

# 1. Título de la Sección
geom_text(aes(x = 0, y = 3.7, label = "Sources and notes"), 
          vjust = 0.7, hjust = 0, lineheight = 0.2, fontface = "bold", size = 6,
          family = "Barlow") +
  
# 2. Línea de Separación
geom_segment(aes(x = 0, y = 3.41, xend = 1.9, yend = 3.41), 
             linewidth = 0.3, color = "black") + 
  
# 3. Texto de Fuentes y Notas
geom_textbox(aes(x = 0, y = 3.4, label = text_sources_alt), 
             vjust = 1, 
             hjust = 0, 
             halign = 0,          
             lineheight = 0.4,    
             size = 4.5,          
             family = "Barlow",
             box.color = NA,      
             fill = NA,           
             width = unit(12, "cm") 
) +
  
scale_fill_identity() + 
scale_color_identity() + 
coord_cartesian(
  xlim = c(0, 2), 
  ylim = c(0, 4.2),
  clip = "off") +
theme_void() + 
theme(
  plot.background = element_rect(fill = "white", color = NA),
  panel.background = element_rect(fill = "white", color = NA),
  plot.margin = margin(0.0, 0, 0, 0, unit = "cm"),
  plot.title = element_text(margin = margin(0, 0, 0, 0, unit = "pt")))

Error bars

This block teaches the user how to read the uncertainty and significance represented in the forest plot.

explanatory_text_sig <- "Statistical significance is crucial\nfor interpreting
the results.\nIt tells us that the relationship\nbetween hunger and
migration\ndesires is likely a real pattern\nin the population, rather
than\na result of random sampling.\nWhen an error bar does not\ncross the 1.0
vertical line, we\ncan be 95% confident in the\ndirection of the effect."

p_error_bar <- ggplot() +
  annotate("text", x = 0, y = 3.8, label = "Statistical\nSignificance", 
           vjust = 0.7, hjust = 0, lineheight = 0.3, fontface = "bold", 
           size = 6, family = "Barlow") +
  geom_segment(aes(x = 0, y = 3.41, xend = 0.85, yend = 3.41), 
               linewidth = 0.4, color = "black") +
  geom_segment(aes(x = 0.1, y = 3, xend = 0.4, yend = 3), 
               color = "gray20", linewidth = 0.4) +
  geom_segment(aes(x = 0.1, y = 2.93, xend = 0.1, yend = 3.07),
               color = "gray20", linewidth = 0.4) +
  geom_segment(aes(x = 0.4, y = 2.93, xend = 0.4, yend = 3.07),
               color = "gray20", linewidth = 0.4) +
  annotate("text", x = 0.5, y = 3, label = "Significant (95% CI)", 
           hjust = 0, size = 5, family = "Barlow") +
  geom_segment(aes(x = 0.1, y = 2.5, xend = 0.4, yend = 2.5), 
               color = "gray85", linewidth = 0.4) +
    geom_segment(aes(x = 0.1, y = 2.43, xend = 0.1, yend = 2.57),
               color = "gray85", linewidth = 0.4) +
  geom_segment(aes(x = 0.4, y = 2.43, xend = 0.4, yend = 2.57),
               color = "gray85", linewidth = 0.4) +
  annotate("text", x = 0.5, y = 2.5, label = "Not significant", 
           hjust = 0, size = 5, family = "Barlow", color = "gray80") +
  annotate("text", x = 0.1, y = 1.1, label = explanatory_text_sig, 
           hjust = 0, vjust = 0.5, lineheight = 0.24, fontface = "plain", 
           size = 6.3, color = "gray70", family = "Caveat") +
  coord_cartesian(xlim = c(0, 1.2), ylim = c(0, 4.2), clip = "off") +
  theme_void() +
  theme(plot.background = element_rect(fill = "white", color = NA),
        plot.margin = margin(t = 0, r = 0, b = 0, l = 0))

Combination main graph and replication

This final stage uses the patchwork library to merge the primary forest plot with the modular legend components into a single, cohesive scientific visualization.

• Modular Row Integration: The three legend components (p_error_bar, p_legend_remittance_alt, and p_sources_alt) are combined into a single horizontal row. We use specific widths (1.2, 1.5, and 2.3) to provide more horizontal space for the text-heavy “Sources” block, preventing the information from feeling cramped.

• Controlled Spacing: By applying panel.spacing = unit(0.5, "cm") with the & operator, we ensure a consistent and clean “gutter” between the different legend sections. This visual separation helps the reader distinguish between statistical methodology, data mapping, and general notes.

• Vertical Hierarchy: The / operator stacks the main chart on top of the legends. We assigned a heights ratio of 5:2, prioritizing the forest plot while ensuring the bottom section is large enough to be easily readable without distracting from the main results.

• Canvas Optimization: The final plot_annotation sets a generous right margin. This is a strategic “safety zone” that prevents the long descriptive text of the sources from being clipped during the rendering process, ensuring a professional finish for web or print publication.

Improved version

p_combined_alt <- alt_main + theme(
    plot.margin = margin(t = 5, r = 5, b = 15, l = 5)
)

library(patchwork)

bottom_row_final_alt <- (p_error_bar | p_legend_remittance_alt | p_sources_alt) +
  plot_layout(widths = c(1.3, 1.5, 2.3)) &
  theme(
    plot.margin = margin(0, 0, 0, 0) 
  )

final_plot_alt <- p_combined_alt / bottom_row_final_alt

final_plot_alt <- final_plot_alt +
  plot_layout(heights = c(5, 2)) +
  plot_annotation(theme = theme(
    plot.margin = margin(t = 10, r = 5, b = 10, l = 5),
    plot.background = element_rect(fill = "white", color = NA)
  ))

Conclusions

1. Statistical Findings: A Contextual Relationship

The analysis confirms that hunger is not a uniform driver of migration. While it acts as a strong “push factor” in countries like The Gambia or Lesotho, its effect is negligible or reversed in others like Tanzania. The inclusion of remittances suggests that a national culture of migration and external financial support often makes emigration a more viable response to food insecurity.

2. Technical Replication: Precision and Micro-design

Replicating the original visualization was a graphic engineering challenge that demanded meticulous attention to details that often go unnoticed:

• Legend Composition: The original legend was not an automatic ggplot2 element but a complex design piece. To replicate it, we had to create four independent small plots (mini-graphs) that were then carefully integrated into the main canvas. This allowed us to represent custom size and color scales exactly as in the source.

• Readability through Shadows: We implemented geom_shadowtext to add subtle white halos to the country labels. This ensured text legibility even when overlapping with grid lines or dense data points.

• Axis Color Gradients: The axes were not treated as simple lines, but as color-gradient segments. We manually calculated dozens of individual segments to replicate the chromatic transition that guides the eye toward areas of higher or lower statistical significance.

• Manual Calibration: Achieving an editorial-grade finish required manual coordinate adjustment for almost every label, proving that high visual fidelity requires direct intervention beyond default software parameters.

3. Enhancements: Clarity Through De-noising

The improved “Forest Plot” version offers several advantages over the original scatter plot:

• Reduced Cognitive Load: By removing population weighting and focusing on the Odds Ratio (OR), the primary research question becomes the focal point.

• Visual Integrity: The dual-tone system (dark vs. light gray) immediately separates statistically significant results from noise, allowing for a “filtering” effect at a glance.

• Symmetry and Scale: The standardized log-scale provides a more honest comparison between “more likely” and “less likely” outcomes, ensuring statistical rigor is matched by visual clarity.

Final Reflection

This project demonstrates that effective data communication is a balance between reproducibility and design. While the replication taught us how to handle complex layouts, the enhancement showed that “less is more”: by simplifying the aesthetics and prioritizing uncertainty, we provide a more accessible and honest answer to the impact of hunger on African migration.