Analyzing inequality in access to opportunities

2024-08-27

Social and racial inequalities in access to opportunities

# load libraries
library(aopdata)
library(sf)
#> Warning: package 'sf' was built under R version 4.3.3
library(ggplot2)
#> Warning: package 'ggplot2' was built under R version 4.3.3
library(data.table)
#> Warning: package 'data.table' was built under R version 4.3.3
library(units)
#> Warning: package 'units' was built under R version 4.3.2

Download accessibility data


df <- read_access(city='Curitiba',
                   mode='public_transport',
                   peak = TRUE,
                   year=2019,
                   showProgress = FALSE)

Inequality in access to job opportunities by income decile


ggplot() +
  geom_boxplot(data=subset(df, !is.na(R003)),
               aes(x = factor(R003), y=CMATT60/1000, color=factor(R003))) +
  scale_color_brewer(palette = 'RdBu') +
  labs(title='Distribution of the number of jobs accessible', color="Income\ndecile",
       subtitle='by public transport in less than 60 min. by income decile',
       x='Income decile', y="N. of jobs accessible\n(thousands)") +
  scale_x_discrete(labels=c("D1 Poorest", paste0('D', 2:9), "D10 Wealthiest")) +
  theme_minimal()

Palma ratio


# average access of the wealthiest 10%
avg_access_10p_wealthiest <- df[ R003==10, weighted.mean(x=CMATT60, w=P001, na.rm=T)]

# average access of the poorest 40%
avg_access_40p_poorest <- df[ R003<=4, weighted.mean(x=CMATT60, w=P001, na.rm=T)]

# Palma ratio
palma_ratio <- avg_access_10p_wealthiest / avg_access_40p_poorest                
palma_ratio 
#> [1] 2.609378

This means that:

message( paste0('In less than 60 min. by public transport, the 10% wealthiest population could access on average ', round(palma_ratio,1), ' times more job opportunites than the 40% poorest people') )

Inequality in travel time to closes hospital


# replace Inf travel time with 120
df[, TMISA := fifelse(TMISA==Inf, 120, TMISA)]

# calculate avarage travel time by race
df[, .(average = weighted.mean(x=TMISA, w=P001, na.rm=T),
       white   = weighted.mean(x=TMISA, w=P002, na.rm=T),
       black   = weighted.mean(x=TMISA, w=P003, na.rm=T))]
#>    average    white    black
#>      <num>    <num>    <num>
#> 1:  37.124 35.75722 43.29197

# calculate avarage travel time by income
temp <- df[, .(average = weighted.mean(x=TMISA, w=P001, na.rm=T)), by=R003]
temp <- na.omit(temp)

ggplot() + 
  geom_point(data=temp, aes(y=average, x=factor(R003))) +
  labs(x='Income decile', y='Avg. travel time to\nclosest hospital') +
  theme_minimal()