library(tidyverse)
# datasource is: IHME https://healthdata.org/
population_2010_2019_long <- population_2010_2019%>%
pivot_longer(cols=3:12,names_to="population_10_19",values_to="pop_value")%>%
rename(location=location_name)
df_cause_and_pop <- bubble_df%>%
inner_join(population_2010_2019_long,by="location")%>%
select(1,3:7,10)
# load more libraries-----------------
library(ggfx)
library(gridExtra)
library(ggrepel)
library(extrafont)
options(scipen = 999)
# plotting -------------------------------
p<-ggplot(data=df_cause_and_pop,aes(x=log10((YLL+YLD)/DALY),y=log10(DALY),size=pop_value)) +
geom_jitter(aes(size=pop_value,alpha=0.5),width = 0.15,height = 1) +
geom_text(aes(label=cause),check_overlap = TRUE)+
coord_polar() +
scale_color_identity() +
labs(title="European Countries",
subtitle="DALYs by Cause years 2010-2019",
caption="Viz Federica Gazzelloni ! DataSource: IHME | Monochromo Day 24",
x="proportion of DALY components",
y="DALYs",
size="Population",
alpha="Transparency")+
theme_grey() +
theme(plot.background = element_rect(fill = "black"),
plot.title=element_text(face="bold",size= 25,family="Impact",color="white"),
plot.subtitle=element_text(face="bold",size= 15,color="white",family="Impact"),
plot.caption=element_text(face="bold",size= 8,color="white",family="Impact"),
panel.grid.major = element_line(color="grey"),
axis.text.y = element_text(size= 8,color="white",family="Impact"),
axis.text.x = element_blank(),
legend.background = element_blank(),
legend.text = element_text(size= 8,color="white",family="Impact"),
legend.title = element_text(size= 8,color="white",family="Impact"))
####### SAVING ######################################
ragg::agg_png(here::here("day24_monochromo", "monochrome_day24.png"),
res = 320, width = 14, height = 8, units = "in")
p
dev.off()