Deep-dive · Plot Gallery

Optional companion to Module 1

Optional. A reference for the plots that come up most often, on the same synthetic data. Module 2 covers this properly; this page is here so you can copy something that works today.

library(readr)
library(dplyr)


Attaching package: 'dplyr'

The following objects are masked from 'package:stats':

    filter, lag

The following objects are masked from 'package:base':

    intersect, setdiff, setequal, union

library(ggplot2)

abcd <- read_csv("data/abcd-synthetic.csv") |> 
  mutate(visit = factor(visit, levels = c("baseline", "year_1", "year_2")))

Rows: 580 Columns: 9

── Column specification ────────────────────────────────────────────────────────
Delimiter: ","
chr (4): subject_id, visit, site, sex
dbl (5): age_months, total_problems, cognition_score, screen_time_hrs, sleep...

ℹ Use `spec()` to retrieve the full column specification for this data.
ℹ Specify the column types or set `show_col_types = FALSE` to quiet this message.

Note the factor() line — it fixes the visit order once, so every plot below comes out chronological rather than alphabetical. See the data types deep-dive for why.

The grammar, in one paragraph

Every ggplot is ggplot(data, aes(...)) plus one or more geom_*() layers joined with +. aes() maps columns to visual properties — x, y, colour, fill, size, shape. Anything you set outside aes() is a fixed constant; anything inside varies with the data. That distinction is the single most common source of confusion:

# colour OUTSIDE aes() -- one fixed colour
ggplot(abcd, aes(x = age_months, y = cognition_score)) +
  geom_point(colour = "#62272D", alpha = 0.5)

# colour INSIDE aes() -- varies by column, legend appears
ggplot(abcd, aes(x = age_months, y = cognition_score, colour = sex)) +
  geom_point(alpha = 0.5)

Two scatterplots side by side; the first is all one colour, the second is coloured by sex. — Figure 1: Left: colour fixed. Right: colour mapped to a column.

One continuous variable

ggplot(abcd, aes(x = cognition_score)) +
  geom_histogram(binwidth = 5, fill = "#62272D", colour = "white") +
  labs(x = "Cognition score", y = "Count")

ggplot(abcd, aes(x = cognition_score)) +
  geom_density(fill = "#FDBF6F", alpha = 0.7) +
  labs(x = "Cognition score", y = "Density")

A histogram of cognition scores next to a smoothed density curve of the same data. — Figure 3: Histogram and density for the same variable

binwidth is worth setting deliberately. Leave it out and ggplot2 picks 30 bins and warns you about it; the right value depends on your data and is a judgement call, so try a few.

One continuous, one categorical

ggplot(abcd, aes(x = visit, y = total_problems)) +
  geom_boxplot(fill = "#FDBF6F") +
  labs(x = NULL, y = "Total problems")

ggplot(abcd, aes(x = visit, y = total_problems)) +
  geom_violin(fill = "#DEEBF7") +
  geom_jitter(width = 0.15, alpha = 0.25, size = 0.7) +
  labs(x = NULL, y = "Total problems")

Boxplots of total problems by visit, and violin plots of the same with individual points jittered on top. — Figure 5: Boxplot, and a violin with the raw points overlaid

The right-hand version shows the actual sample size and any clumping, which a boxplot hides. With a few hundred points it is usually the more honest choice.

Two continuous variables

ggplot(abcd, aes(x = age_months, y = cognition_score)) +
  geom_point(alpha = 0.4, colour = "#62272D") +
  geom_smooth(method = "lm", colour = "#FF7F00") +
  labs(x = "Age (months)", y = "Cognition score")

Scatterplot of cognition score against age in months with a fitted straight line and confidence band. — Figure 7: Scatterplot with a linear trend

geom_smooth(method = "lm") draws a straight-line fit with a confidence band. Leave method out and you get a wiggly loess curve instead, which is better for “is there any pattern here at all?” and worse for “is this trend linear?”.

Warning

Drawing a trend line is not the same as testing one. This picture has no p-value and makes no claim. Module 3 is where fitted lines become inference.

Change over time

Two different questions, two different plots.

Group averages — one line, the summary:

abcd |> 
  group_by(visit) |> 
  summarise(
    mean_cog = mean(cognition_score, na.rm = TRUE),
    se       = sd(cognition_score, na.rm = TRUE) / sqrt(sum(!is.na(cognition_score)))
  ) |> 
  ggplot(aes(x = visit, y = mean_cog, group = 1)) +
  geom_line(colour = "#62272D", linewidth = 1) +
  geom_point(size = 3, colour = "#62272D") +
  geom_errorbar(aes(ymin = mean_cog - se, ymax = mean_cog + se), width = 0.1) +
  labs(x = NULL, y = "Mean cognition score")

Line chart with error bars showing mean cognition score increasing across three visits. — Figure 8: Mean cognition score by visit, with standard-error bars

Individual trajectories — one line per participant, the spaghetti plot:

set.seed(1)
some_ids <- sample(unique(abcd$subject_id), 30)

abcd |> 
  filter(subject_id %in% some_ids) |> 
  ggplot(aes(x = visit, y = cognition_score, group = subject_id)) +
  geom_line(alpha = 0.35, colour = "#62272D") +
  geom_point(alpha = 0.5, size = 1) +
  stat_summary(aes(group = 1), fun = mean, geom = "line",
               colour = "#FF7F00", linewidth = 1.4) +
  labs(x = NULL, y = "Cognition score")

Many faint lines, one per participant, each connecting their cognition scores across visits, with a thick average line on top. — Figure 9: Individual trajectories for 30 participants

This is the plot that makes longitudinal data feel real. The orange average rises gently; individual children go up, down and sideways. Any analysis that treats these 90 points as independent observations is wrong, and that is precisely the problem Module 3 takes on.

Note group = subject_id: it tells ggplot2 which points belong to the same person. Without it you get one meaningless zigzag.

Small multiples

facet_wrap() splits one plot into a panel per group. Usually clearer than cramming four colours into one panel:

ggplot(abcd, aes(x = age_months, y = cognition_score)) +
  geom_point(alpha = 0.4, size = 0.8, colour = "#62272D") +
  geom_smooth(method = "lm", se = FALSE, colour = "#FF7F00") +
  facet_wrap(~ site) +
  labs(x = "Age (months)", y = "Cognition score")

~ site is a formula — read ~ as “by”. The same notation appears throughout R’s modelling functions in Module 3.

Counts

ggplot(abcd, aes(x = site, fill = sex)) +
  geom_bar() +
  labs(x = NULL, y = "Participant-visits")

ggplot(abcd, aes(x = site, fill = sex)) +
  geom_bar(position = "dodge") +
  labs(x = NULL, y = "Participant-visits")

Stacked bars are good for totals, bad for comparing the parts. position = "dodge" is better when the comparison between groups is the point.

Making it presentable

A plot for yourself needs nothing. A plot for a poster needs labels, a theme, and a deliberate size:

ggplot(abcd, aes(x = visit, y = total_problems, fill = visit)) +
  geom_boxplot(show.legend = FALSE, outlier.alpha = 0.3) +
  scale_fill_manual(values = c("#DEEBF7", "#FDBF6F", "#62272D")) +
  labs(
    title    = "Reported behaviour problems decline modestly across visits",
    subtitle = "Synthetic ABCD-style data, 200 participants",
    x        = NULL,
    y        = "Total problems",
    caption  = "Synthetic data — for teaching only"
  ) +
  theme_minimal(base_size = 12) +
  theme(plot.title = element_text(face = "bold"))

A polished boxplot of total problems by visit with title, subtitle, caption and a clean minimal theme. — Figure 13: The same data, tidied up for presentation

Then save it at a size and resolution a journal will accept:

ggsave("output/problems-by-visit.png", width = 7, height = 4.5, dpi = 300)

ggsave() saves the last plot drawn. Set width and height in inches — resizing a saved PNG afterwards makes the text blurry, whereas re-saving at the right size does not.

Back to Module 1, or on to Module 2 where this becomes the main event.