Quantitative visualisation

Day 2 · 09:30

ggplot2 is a grammar, not a gallery. Once you can name the data, the aesthetics and the geom, you can build any board chart — and you can finally stop shipping bars that hide the distribution.

This session is the commercial rewrite of a lesson that used to involve haggis and porridge. You will use a breakfast panel from a national grocer: households that buy oats versus instant cereal, in modern trade versus the open market. The question management asked was *which breakfast drives higher category spend*. The data will refuse to answer that until you look at the distribution.

The grammar

library(tidyverse)
 
# data  %>%  ggplot(aes(...))  +  geom_*()  +  labs()
 
breakfast <- read_csv("breakfast-panel.csv")
 
breakfast %>%
  ggplot(aes(age, monthly_spend_ngn)) +
  geom_point(alpha = 0.45) +
  geom_smooth(method = "lm") +
  labs(
    title = "Age and category spend",
    x = "Age",
    y = "Monthly spend (NGN)"
  )

• data — a tidy table. One row, one household.
• aes() — which variables become position, colour, size.
• geom — what to draw: point, smooth, violin, col, tile.
• Geoms are additive. A line of best fit on top of points is two geoms, not a new chart type.

Colour is a variable

breakfast %>%
  ggplot(aes(age, monthly_spend_ngn, colour = group)) +
  geom_point(alpha = 0.5) +
  geom_smooth(method = "lm", se = FALSE) +
  scale_colour_manual(values = c(oats = "#0B1F3A", instant = "#B8963E")) +
  labs(colour = "Breakfast group")

The instant line is almost flat. The oats line is not. A single colour-blind scatter would have invited a single slope — and a single, wrong, national promotion.

Ban the bar

A bar of means with a ±1 SD error bar is the default in too many packs. It hides skew, hides multimodality, and invents a symmetry the data do not have. When EXCO asks *‘are the customers clustered or spread?’* the bar cannot answer.

breakfast %>%
  ggplot(aes(group, monthly_spend_ngn, fill = trade_channel)) +
  geom_violin(alpha = 0.45, position = position_dodge(width = 0.9)) +
  geom_boxplot(width = 0.15, position = position_dodge(width = 0.9),
               outlier.size = 0.8) +
  scale_fill_manual(values = c(modern_trade = "#0B1F3A",
                               open_market  = "#B8963E")) +
  labs(
    title = "Spend by breakfast group and trade channel",
    y = "Monthly spend (NGN)",
    fill = "Channel"
  )

Reading the interaction: instant cereal spend barely moves between modern trade and the open market. Oats spend does. A campaign that treats ‘breakfast’ as one number will over-invest in a group whose variance is a channel story, not a brand story.

A banking twin

The same grammar applies to nps-by-channel.csv. Mobile is not ‘better’ because its mean is higher. Check whether the left tail — the detractors — is fatter on USSD, which is where a regulator will look.

Labels, themes, files

last_plot() +
  labs(
    title = "NPS by channel",
    subtitle = "Distributions, last 90 days — not a league table",
    caption = "Source: Northridge Bank voice file, 2026 Q2 extract"
  ) +
  theme_minimal(base_family = "Source Sans 3")
 
ggsave("nps-channel.png", width = 9, height = 5.5, units = "in")