This document gives a few suggestions to analyse a dataset composed by a numeric variable measured on groups and subgroups, with several observations for each combinations.

The dataset used as an example describes the amount that restaurant staff receive in tips based on various indicators like the client sex, the day of the week and if the client smokes or not.

Data come from the Seaborn Python library. A clean .csv version is available in the data-to-viz.com github repository.

# Libraries
library(tidyverse)
library(hrbrthemes)
library(kableExtra)
options(knitr.table.format = "html")
library(viridis)
library(ggrepel)
library(plotly)

# Load dataset from github
data <- read.table("https://raw.githubusercontent.com/holtzy/data_to_viz/master/Example_dataset/10_OneNumSevCatSubgroupsSevObs.csv", header=T, sep=",") %>%
  mutate(tip = round(tip/total_bill*100, 1))

# show data
data %>% head(6) %>% select(tip, sex, day, smoker) %>% kable() %>%
  kable_styling(bootstrap_options = "striped", full_width = F)

tip	sex	day	smoker
5.9	Female	Sun	No
16.1	Male	Sun	No
16.7	Male	Sun	No
14.0	Male	Sun	No
14.7	Female	Sun	No
18.6	Male	Sun	No

Grouped boxplot

The most common way to represent that kind of dataset is probably the grouped boxplot. Each combination of group is represented by a box. This allows to quickly understand what is the distribution of the numeric variable for each combination.

Here, it looks like there is not much difference in tip values from one day to the other in average, except a slight increase on sunday. Moreover, it looks like females tend to tip more than males on friday. Note that individual data points are presented using jittering, what allows to detect more particular pattern and assess the sample size of each group.

# Counts the number of value per group and subgroup
counts = data %>%
  group_by(day, sex) %>%
  summarize(
    n=n(),
    median=median(tip)
  )

# Grouped
data %>%
  mutate(day = fct_reorder(day, tip)) %>%
  mutate(day = factor(day, levels=c("Thur", "Fri", "Sat", "Sun"))) %>%
  ggplot(aes(fill=sex, y=tip, x=day)) +
    geom_boxplot(position=position_dodge2(preserve = "total"), alpha=0.5, outlier.colour="transparent", varwidth = TRUE) +
    geom_point(color="grey", size=1, width=0.1, position=position_jitterdodge() , alpha=0.4) +
    scale_fill_viridis(discrete=T, name="") +
    geom_text(data=counts, aes(label=paste0("n: ",n), y=median-2), position=position_dodge(1), hjust=0.5) +
    theme_ipsum()  +
    xlab("") +
    ylab("Tip (%)") +
    ylim(0,35)

In the above chart categories are grouped by day. It is possible to build the same kind of visualization, grouping by Sex instead. In the first case, it is easy to compare the behavior or each sex day by day. On the second one, the idea is more to compare the difference by day for each sex.

# Grouped
data %>%
  mutate(day = fct_reorder(day, tip)) %>%
  mutate(day = factor(day, levels=c("Thur", "Fri", "Sat", "Sun"))) %>%
  ggplot(aes(fill=day, y=tip, x=sex)) +
    geom_boxplot(position="dodge", alpha=0.5, outlier.colour="transparent") +
    geom_point(color="grey", size=1, width=0.1, position=position_jitterdodge() , alpha=0.4) +
    scale_fill_viridis(discrete=T, name="") +
    theme_ipsum()  +
    xlab("") +
    ylab("Tip (%)") +
    ylim(0,40)

Violin

The violin plot can be used exactly in the same situation that the boxplot. Using the same grouping technique, we got here a grouped violin plot.

# Grouped
data %>%
  mutate(day = fct_reorder(day, tip)) %>%
  mutate(day = factor(day, levels=c("Thur", "Fri", "Sat", "Sun"))) %>%
  ggplot(aes(fill=sex, y=tip, x=day)) +
    geom_violin(position="dodge", alpha=0.5, outlier.colour="transparent") +
    scale_fill_viridis(discrete=T, name="") +
    theme_ipsum()  +
    xlab("") +
    ylab("Tip (%)") +
    ylim(0,40)

Histogram

Small multiple is a powerful technique that can be used with that kind of data. Here, each combination is represented in a part of the grid, allowing a quick comparison between them. Histograms are used here, but the same result could be obtained with density plot.

# Grouped
data %>%
  mutate(day = fct_reorder(day, tip)) %>%
  mutate(day = factor(day, levels=c("Thur", "Fri", "Sat", "Sun"))) %>%
  ggplot(aes(x=tip)) +
    geom_histogram(bins=20, fill="#69b3a2", color="white") +
    facet_grid(sex~time) +
    theme_ipsum()  +
    xlab("") +
    ylab("Tip")

Going further

You can learn more about each type of graphic presented in this story in the dedicated sections. Click the icon below:

Grouped boxplot

Violin

Histogram

Going further

Dataviz decision tree