In this chapter, we will focus on multivariate categorical data. Here, it is noteworthy that multivariate plot is not the same as multiple variable plot, where the former is used for analysis with multiple outcomes.
Bar chats are used to display the frequency of multidimensional categorical variables. In the next few plots you will be shown different kinds of bar charts.
library(tidyverse)
cases <- read.csv("data/icecream.csv") |>
mutate(Age = fct_relevel(Age, "young"))
icecreamcolors <- c("#ff99ff", "#cc9966") # pink, coffee
ggplot(cases, aes(x = Age, fill = Favorite)) +
geom_bar() + scale_fill_manual(values = icecreamcolors)
Use position = "dodge" to create grouped bar chart
ggplot(cases, aes(x = Age, fill = Favorite)) +
geom_bar(position = "dodge") +
scale_fill_manual(values = icecreamcolors)
ggplot(cases, aes(x = Age)) +
geom_bar(position = "dodge") +
facet_wrap(~Favorite)
counts3 <- cases |>
group_by(Age, Favorite, Music) |>
summarize(Freq = n()) |>
ungroup() |>
complete(Age, Favorite, Music, fill = list(Freq = 0))
ggplot(counts3, aes(x = Favorite, y = Freq, fill = Music)) +
geom_col(position = "dodge") +
facet_wrap(~Age)
In this section, we would like to show how to use chi-square test to check the independence between two features.
We will use the following example to answer: Are older Americans more interested in local news than younger Americans? The dataset is collected from here.
local <- data.frame(Age = c("18-29", "30-49", "50-64", "65+"),
Freq = c(2851, 9967, 11163, 10911)) |>
mutate(Followers = round(Freq*c(.15, .28, .38, .42)),
Nonfollowers = Freq - Followers) |>
select(-Freq)
knitr::kable(local[,1:2])| Age | Followers |
|---|---|
| 18-29 | 428 |
| 30-49 | 2791 |
| 50-64 | 4242 |
| 65+ | 4583 |
The chi-square hypothesis is set to be:
Null hypothesis: Age and tendency to follow local news are independent
Alternative hypothesis: Age and tendence to follow local news are NOT independent
localmat <- as.matrix(local[,2:3])
rownames(localmat) <- local$Age
X <- chisq.test(localmat, correct = FALSE)
X$observed Followers Nonfollowers
18-29 428 2423
30-49 2791 7176
50-64 4242 6921
65+ 4583 6328X$expected Followers Nonfollowers
18-29 984.1065 1866.893
30-49 3440.4032 6526.597
50-64 3853.2378 7309.762
65+ 3766.2526 7144.747X
Pearson's Chi-squared test
data: localmat
X-squared = 997.48, df = 3, p-value < 2.2e-16We compare observed to expected and then the p-value tells that age and tendency are independent features. We are good to move on to next stage on mosaic plots.
Mosaic plots are used for visualizing data from two or more qualitative variables to show their proportions or associations.
library(grid)
icecream <- read.csv("data/MusicIcecream.csv") |>
mutate(Age = fct_relevel(Age, "young"))
icecreamcolors <- c("#ff99ff", "#cc9966")
counts2 <- icecream |>
group_by(Age, Favorite) |>
summarize(Freq = sum(Freq))
vcd::mosaic(~Age, direction = "v", counts2)
vcd::mosaic(Favorite ~ Age, counts2, direction = c("v", "h"),
highlighting_fill = icecreamcolors)
Here’s some criteria of best practice of mosaic plots :
Dependent variables is split last and split horizontally
Fill is set to dependent variable
Other variables are split vertically
Most important level of dependent variable is closest to the x-axis and darkest (or most noticable shade)
vcd::mosaic(Favorite ~ Age + Music, counts3,
direction = c("v", "v", "h"),
highlighting_fill = icecreamcolors)
Use pairs method to plot a matrix of pairwise mosaic plots for class table:
pairs(table(cases[,2:4]), highlighting = 2)
Spine plot is a mosaic plot with straight, parallel cuts in one dimension (“spines”) and only one variable cutting in the other direction.
library(vcdExtra)
library(forcats)
foodorder <- Alligator |> group_by(food) |> summarize(Freq = sum(count)) |>
arrange(Freq) |> pull(food)
ally <- Alligator |>
rename(Freq = count) |>
mutate(size = fct_relevel(size, "small"),
food = factor(food, levels = foodorder),
food = fct_relevel(food, "other"))
vcd::mosaic(food ~ sex + size, ally,
direction = c("v", "v", "h"),
highlighting_fill= RColorBrewer::brewer.pal(5, "Accent"))
Treemap is a filled rectangular plot representing hierarchical data (fill color does not necessarily represent frequency count)
library(treemap)
data(GNI2014)
treemap::treemap(GNI2014,
index=c("continent", "iso3"),
vSize="population",
vColor="GNI",
type="value",
format.legend = list(scientific = FALSE, big.mark = " "))
This type of chart works well with likert data, or any ordinal data with categories that span two opposing poles. The code below uses the likert() function from the HH package.
library(HH)
gdata <- read_csv("data/gender.csv")
HH::likert(Group~., gdata, positive.order = TRUE,
col=likertColorBrewer(3, ReferenceZero = NULL,
BrewerPaletteName = "BrBG"),
main = "% saying the country __ when \n it comes to giving women equal rights with men",
xlab = "percent", ylab = "")
Use | to condition (facet) on factor levels
gdata$Section <- c("Overall", "Gender", "Gender", "Party", "Party")
gdata <- gdata |> dplyr::select(Section, Group, everything())
# sort facets manually
gdata <- gdata |> mutate(Section = factor(Section,
levels = c("Party", "Gender", "Overall")))
likert(Group ~ . | Section,
data = gdata,
scales = list(y = list(relation = "free")), # equivalent to scales = "free_y"
layout = c(1, 3), # controls position of subplots
positive.order = TRUE,
col=likertColorBrewer(3, ReferenceZero = NULL,
BrewerPaletteName = "BrBG"),
main = "% saying the country __ when \n it comes to giving women equal rights with men",
xlab = "percent",
ylab = NULL)
Reference: R. Heiberger and N. Robbins, Design of Diverging Stacked Bar Charts for Likert Scales and Other Applications
Alluvial diagrams are usually used to represent the flow changes in network structure over time or between different levels.
The following plot shows the essential components of alluvial plots used in the naming schemes and documentation (axis, alluvium, stratum, lode):
library(ggalluvial)
df2 <- data.frame(Class1 = c("Stats", "Math", "Stats", "Math", "Stats", "Math", "Stats", "Math"),
Class2 = c("French", "French", "Art", "Art", "French", "French", "Art", "Art"),
Class3 = c("Gym", "Gym", "Gym", "Gym", "Lunch", "Lunch", "Lunch", "Lunch"),
Freq = c(20, 3, 40, 5, 10, 2, 5, 15))
ggplot(df2, aes(axis1 = Class1, axis2 = Class2, axis3 = Class3, y = Freq)) +
geom_alluvium(color='black') +
geom_stratum() +
geom_text(stat = "stratum", aes(label = paste(after_stat(stratum), "\n", after_stat(count)))) +
scale_x_discrete(limits = c("Class1", "Class2", "Class3"))
You can choose to color the alluvium by different variables, for example, the first variable Class1 here:
ggplot(df2, aes(axis1 = Class1, axis2 = Class2, axis3 = Class3, y = Freq)) +
geom_alluvium(aes(fill = Class1), width = 1/12) +
geom_stratum() +
geom_text(stat = "stratum", aes(label = paste(after_stat(stratum), "\n", after_stat(count)))) +
scale_x_discrete(limits = c("Class1", "Class2", "Class3"))
Another way of plotting alluvial diagrams is using geom_flow rather than geom_alluvium:
ggplot(df2, aes(axis1 = Class1, axis2 = Class2, axis3 = Class3, y = Freq)) +
geom_flow(aes(fill = Class1), width = 1/12) +
geom_stratum() +
geom_text(stat = "stratum", aes(label = paste(after_stat(stratum), "\n", after_stat(count)))) +
scale_x_discrete(limits = c("Class1", "Class2", "Class3"))
After we use geom_flow, all Math students learning Art came together, which is also the same as Stats students. It makes the graph much clearer than geom_alluvium since there is less cross alluviums between each axises.
Besides what have been systematically introduced in Chapter 9.2 Heatmaps, this part demonstrated a special case of heat map when both x and y are categorical. Here the heat map can been seen as a clustered bar chart and a pre-defined theme is used to show the dense more clearly.
library(vcdExtra)
library(dplyr)
theme_heat <- theme_classic() +
theme(axis.line = element_blank(),
axis.ticks = element_blank())
orderedclasses <- c("Farm", "LoM", "UpM", "LoNM", "UpNM")
mydata <- Yamaguchi87
mydata$Son <- factor(mydata$Son, levels = orderedclasses)
mydata$Father <- factor(mydata$Father,
levels = orderedclasses)
mydata3 <- mydata |> group_by(Country, Father) |>
mutate(Total = sum(Freq)) |> ungroup()
ggplot(mydata3, aes(x = Father, y = Son)) +
geom_tile(aes(fill = (Freq/Total)), color = "white") +
coord_fixed() +
scale_fill_gradient2(low = "black", mid = "white",
high = "red", midpoint = .2) +
facet_wrap(~Country) + theme_heat