Waking Up the Golden Dawn
Does Exposure to the Refugees Increase Support for Extreme-Right Parties?
Theoretical Framework
Causal Framework (DAG)
Show the code
# ------------------------------------------------------------
# Goal: Draw a simple Directed Acyclic Graph (DAG) that represents
# a theory of how exposure to refugees affects Golden Dawn (GD) vote share.
# This DAG uses ggdag + dagitty + ggplot2.
#
# Key idea: We draw *nodes* (circles) for variables and *arrows* for
# hypothesized causal effects. We'll also show how to make
# arrows shorter so they don't touch the nodes.
# ------------------------------------------------------------
# --- 0) Playground: tweak these safely -------------------------------
node_size <- 25 # how big each node (circle) is
node_alpha <- 0.5 # transparency of node fill (0 = invisible, 1 = solid)
arrowhead_pt <- 10 # arrowhead size in points (this is the *tip* triangle)
trim <- 0.20 # how much of the arrow *shaft* to chop off from BOTH ends
# e.g., 0.20 = remove 20% at the start AND 20% at the end
# so the arrow in the middle is 60% of the original length
# --- 1) Load required packages ----------------------------------------------
# ggdag → helper functions to draw DAGs in ggplot2
# dagitty → defines DAG objects and relationships between variables
# ggplot2 → plotting system used to visualize the DAG
library(ggdag)
library(dagitty)
library(ggplot2)
# --- 2) Define the DAG structure --------------------------------------------
# dagify() creates a DAG object where you specify arrows (causal paths).
# The notation A ~ B means "A is caused by B".
# Here:
# GDvote ← threat (Golden Dawn vote share caused by perceived threat)
# threat ← exposure (Threat is caused by exposure to refugees)
# exposure ← distance (Exposure caused by proximity to Turkey)
# Labels: provide readable names for the DAG nodes (with line breaks).
# Coords: x and y positions of each node (for nice plotting layout).
theory_dag <- dagify(
GDvote ~ threat, # DV (Golden Dawn vote share) depends on threat
threat ~ exposure, # Threat depends on exposure
exposure ~ distance, # Exposure depends on distance
labels = c(
GDvote = "Golden Dawn\nVote Share (DV)",
threat = "Perceived Threat\n(Mediator)",
exposure = "Refugee Arrivals",
distance = "Proximity to Turkey"
),
coords = list(
x = c(distance = 0, exposure = 1, threat = 2, GDvote = 3),
y = c(distance = 1, exposure = 1, threat = 1, GDvote = 1)
)
)
# --- 3) Convert DAG into a data frame for ggplot ----------------------------
# tidy_dagitty() extracts coordinates, labels, and edges for plotting.
# as.data.frame() makes it into a tibble/data frame usable by ggplot.
dag_df <- as.data.frame(tidy_dagitty(theory_dag, layout = "auto"))
# Extract plotting limits (so the nodes don’t touch the plot borders).
min_x <- min(dag_df$x); max_x <- max(dag_df$x)
min_y <- min(dag_df$y); max_y <- max(dag_df$y)
error <- 0.2 # small padding around the plot
# --- 3.5) Make arrow *segments* shorter and floating between nodes ----------
# Important: In ggplot, the "arrow()" only controls the arrowhead (the little
# triangle tip). It does NOT control the length of the shaft. Shaft length is
# determined by the start (x, y) and end (xend, yend) coordinates.
#
# To make arrow shafts shorter AND not start at the node centers, we:
# 1) Find all edge rows (they have xend/yend values).
# 2) Move the start point forward by 'trim' proportion along the edge.
# 3) Move the end point backward by 'trim' proportion along the edge.
# The result: the arrow appears between nodes without touching them.
edges_df <- subset(dag_df, !is.na(xend) & !is.na(yend))
# Safety checks: ensure trim is between 0 and 0.49
# (If trim >= 0.5 the start could pass the end!)
trim <- max(min(trim, 0.49), 0.00)
# Compute trimmed start (xstart_trim, ystart_trim) and end (xend_trim, yend_trim)
edges_df$xstart_trim <- edges_df$x + trim * (edges_df$xend - edges_df$x)
edges_df$ystart_trim <- edges_df$y + trim * (edges_df$yend - edges_df$y)
edges_df$xend_trim <- edges_df$xend - trim * (edges_df$xend - edges_df$x)
edges_df$yend_trim <- edges_df$yend - trim * (edges_df$yend - edges_df$y)
# --- 4) Plot the DAG --------------------------------------------------------
# Each node is a big semi-transparent point with a label on top.
# Edges (arrows) are drawn with geom_dag_edges_arc using the *trimmed* endpoints.
ggplot(data = dag_df) +
# Draw big transparent points for each node
geom_dag_point(aes(x = x, y = y), size = node_size, alpha = node_alpha) +
# Add text labels (only once per node → !duplicated(label))
geom_label(
data = subset(dag_df, !duplicated(label)),
aes(x = x, y = y, label = label),
fill = alpha("white", 0.8)
) +
# Draw arrows between nodes (edges of the DAG)
# NOTE: arrow(length = ...) controls ONLY the arrowhead size (the tip).
# The shaft length is controlled by the start/end coordinates we just trimmed.
geom_dag_edges_arc(
data = edges_df,
aes(x = xstart_trim, y = ystart_trim, xend = xend_trim, yend = yend_trim),
curvature = 0.0,
arrow = grid::arrow(length = grid::unit(arrowhead_pt, "pt"), type = "closed")
) +
# Set x- and y-limits to keep everything inside the plot
coord_sf(
xlim = c(min_x - error, max_x + error),
ylim = c(min_y - error, max_y + error)
) +
# Minimal background → removes axes, gridlines, etc.
theme_void()
Figure 2
Methodology
Parallel Trends Check
Show the code
# ------------------------------------------------------------
# Goal: Prepare panel features (trends, lags, diffs, dummies),
# build grouped summaries with CIs, and plot GD vote % over time
# by treatment status.
# ------------------------------------------------------------
# --- 0) Make sure 'year' is numeric -----------------------------------------
# Many imported datasets store years as character or labelled types.
# Converting to integer avoids problems when sorting, lagging, etc.
datamuni$year <- as.integer(datamuni$year)
# --- 1) Municipality-specific linear trends (Model 2, etc.) -----------------
# Some specifications include a separate linear trend for each municipality.
# Your model expects columns named s2trend, s3trend, ..., s94trend explicitly.
# Each trend is "year" for that municipality, and 0 for all others.
# NOTE:
# - This starts at m = 2 to match your model’s expected naming.
# - Requires 'datamuni$muni' to be an integer ID in [1..94].
for (m in 2:94) {
trend_col <- paste0("s", m, "trend")
datamuni[[trend_col]] <- ifelse(datamuni$muni == m, datamuni$year, 0L)
}
# --- 2) Lags (only for GD in percentage terms) -------------------------------
# We create:
# - gdper : Golden Dawn vote share in percent
# - gdlagper : 1-period lag of GD vote share (in percent)
# - gdlag3per : 3-period lag (in percent)
# Ordering by muni, then year, ensures lags are computed in time order.
# IMPORTANT:
# - Lags produce NA in the first available years per municipality.
# - 'gdvote' is assumed to be a proportion (0–1). Multiply by 100 for %.
library(dplyr)
datamuni <- datamuni %>%
arrange(muni, year) %>%
group_by(muni) %>%
mutate(
gdper = gdvote * 100, # main DV in percent
gdlagper = dplyr::lag(gdvote, 1) * 100, # 1-year lag, in percent
gdlag3per = dplyr::lag(gdvote, 3) * 100 # 3-year lag, in percent
) %>%
ungroup()
# --- 3) Differences for Reduced Form / Placebo figures -----------------------
# First difference of GD % (current minus 1-year lag), used in fig4a.
# A "placebo" style difference using 1-year-lag minus 3-year-lag, used in fig4b.
# NOTE: These will be NA where required lags are missing.
datamuni <- datamuni %>%
mutate(
gdperdif = gdper - gdlagper, # ΔGD% (t − t−1)
gdperdif3 = gdlagper - gdlag3per # ΔGD% (t−1 − t−3)
)
# --- 4) Year dummies + interactions (Table S5 / Fig. S4) ---------------------
# Create indicator (dummy) variables for specific years and interact with logdist.
# Here, y1..y4 correspond to years 2012, 2013, 2015, 2016 respectively.
# The interactions (y*dist) let the effect of distance vary by those years.
# Assumes 'logdist' already exists (e.g., log(distance to Turkey)).
datamuni <- datamuni %>%
mutate(
y1 = as.integer(year == 2012),
y2 = as.integer(year == 2013),
y3 = as.integer(year == 2015),
y4 = as.integer(year == 2016),
y1dist = y1 * logdist,
y2dist = y2 * logdist,
y3dist = y3 * logdist,
y4dist = y4 * logdist
)
# --- 5) Other percent levels for appendix summaries --------------------------
# Convert other commonly used variables to percent levels (no lags/diffs here).
# 'ndvote' and 'turnout' are assumed to be proportions (0–1).
datamuni <- datamuni %>%
mutate(
ndper = ndvote * 100, # New Democracy in %
turnper = turnout * 100 # Turnout in %
)
# --- 6) Grouped means, SDs, and 95% CIs by treatment & year ------------------
# We compute:
# - mean GD% (gdper_mean)
# - SD of GD% (gdper_sd)
# - sample size (count_n)
# - standard error (se = sd/sqrt(n))
# - 95% CI ≈ mean ± 1.96 * se (normal approx)
# NOTE:
# - If your data contain NA values, consider mean(..., na.rm = TRUE) / sd(..., na.rm = TRUE).
# - 'treatment_group' is assumed to be coded (e.g., 0 = Control, 1 = Treated).
df_muni_avg2 <- datamuni %>%
mutate(gdper = gdvote * 100) %>%
group_by(treatment_group, year) %>%
summarize(
gdper_mean = mean(gdper), # add na.rm = TRUE if needed
gdper_sd = sd(gdper), # add na.rm = TRUE if needed
count_n = n(),
.groups = "drop"
) %>%
mutate(
se_gdper = gdper_sd / sqrt(count_n),
gdcilo = gdper_mean - 1.96 * se_gdper, # lower 95% bound
gdcihi = gdper_mean + 1.96 * se_gdper # upper 95% bound
)
# --- 7) Plot: GD% over time by treatment group with 95% CIs ------------------
# - Lines/points show average GD% over time within each group.
# - Error bars visualize 95% confidence intervals.
# - Colors: blue = Control, red = Treated (per your manual scale).
# - Y-axis fixed to [2, 11] with 1-pt ticks for readability (match your figure).
fig2a <- ggplot(
df_muni_avg2,
aes(x = year, y = gdper_mean,
color = as.factor(treatment_group),
group = treatment_group)
) +
geom_line(size = 1, position = position_dodge(width = 0.2)) +
geom_point(size = 2.5, position = position_dodge(width = 0.2)) +
geom_errorbar(aes(ymin = gdcilo, ymax = gdcihi),
width = 0.2,
position = position_dodge(width = 0.2)) +
scale_color_manual(
values = c("blue", "red"),
labels = c("Control", "Treated"),
name = "Group"
) +
theme_bw() +
scale_y_continuous(
limits = c(2, 11),
breaks = seq(2, 11, by = 1)
) +
theme(
legend.position = c(0.00, 1.00), # place legend at top-left inside plot
legend.justification = c("left", "top"),
legend.background = element_rect(fill = "white", color = "black")
)
# --- (Optional) Save the figure to disk --------------------------------------
# ggsave(here::here("figures", "gd_trends_by_treatment.png"),
# fig2a, width = 7, height = 4.5, dpi = 300)
# Display the figure
print(fig2a)
Figure 3
Findings
Robustness: Distance to Turkey
Show the code
# ------------------------------------------------------------
# Goal: Plot reduced-form and placebo relationships between
# logged distance to Turkey and changes in Golden Dawn vote share.
# ------------------------------------------------------------
# --- 1) Subset the data to 2016 only ----------------------------------------
# We focus on 2016 (the election of interest).
# 'datamuni' must already contain:
# - logdist : logged distance to Turkey
# - gdperdif : ΔGD% (current − lagged, Jan–Sep 2015 → main reduced form)
# - gdperdif3 : placebo difference (lagged − 3-lagged)
red <- subset(datamuni, year == 2016)
# --- 2) Reduced Form Figure -------------------------------------------------
# Scatterplot: x = log distance, y = change in GD vote share
# Add linear fit (stat_smooth with method = "lm")
# Limit axes so the figure matches published version (y: −5..5, x: 0.2..6.3)
fig4a <- ggplot(red, aes(x = logdist, y = gdperdif)) +
geom_point(shape = 1) + # open circles
ylim(-5, 5) + # y-axis limits
xlim(0.2, 6.3) + # x-axis limits
stat_smooth(method = "lm") + # add OLS fit line
xlab("Logged Distance in Km") + # x-axis label
ylab("Change in GD vote: Jan–Sep 2015") + # y-axis label
theme(legend.position = "none") + # no legend needed
ggtitle("Reduced Form") + # title
theme_bw() # clean theme
# --- 3) Placebo Figure ------------------------------------------------------
# Same plot setup, but outcome is gdperdif3 (placebo difference).
# The placebo test checks whether “fake” changes line up with distance.
fig4b <- ggplot(red, aes(x = logdist, y = gdperdif3)) +
geom_point(shape = 1) +
ylim(-5, 5) +
xlim(0.2, 6.3) +
stat_smooth(method = "lm") +
xlab("Logged Distance in Km") +
ylab("Change in GD vote: Jan–Sep 2015") +
theme(legend.position = "none") +
ggtitle("Placebo") +
theme_bw()
# --- 4) Arrange the two figures side by side --------------------------------
# ggarrange() puts both figures into one row with two columns.
ggarrange(fig4a, fig4b, ncol = 2)
Figure 4