Data Analysis for LEL - Week 8

Advanced plotting

Stefano Coretta

University of Edinburgh

Mosaic plot

glot_status |> 
  filter(status != "extinct") |> 
  droplevels() |> 
  ggplot() +
  # from the ggmosaic package
  geom_mosaic(
    aes(product(Macroarea), fill = status),
    alpha = 1) +
  scale_fill_brewer(palette = "BuPu") +
  theme_dark() +
  theme(legend.position = "bottom")

Mosaic plot

Line plot

formants %>%
  filter(c2 %in% c("t", "k")) |>
  filter(vowel %in% c("a", "o")) |> 
  ggplot(aes(time, value, group = id, colour = formant)) +
  geom_line(alpha = 0.8) +
  facet_grid(cols = vars(c2), rows = vars(vowel)) +
  scale_color_brewer(type = "qual") +
  labs(
    x = "Relative time", y = "Formant value (Hz)",
    title = "Formant trajectories of Italian /a, o/ followed by /k/ or /t/"
  )

Line plot

Line plot with points

formants %>%
  filter(c2 %in% c("t", "k")) |>
  filter(vowel %in% c("a", "o")) |> 
  ggplot(aes(time, value, group = id,
             colour = formant)) +
  geom_line(alpha = 0.25) +
  geom_point(alpha = 0.5) +
  facet_grid(cols = vars(c2), rows = vars(vowel)) +
  scale_color_brewer(type = "qual") +
  labs(
    x = "Relative time", y = "Formant value (Hz)",
    title = "Formant trajectories of Italian /a, o/ followed by /k/ or /t/"
  )

Line plot with points

Connected dots plot

gestures |> 
  filter(gesture != "ho_gv", count < 30) |> 
  group_by(dyad, background, months, gesture) |> 
  summarise(
    count_tot = sum(count),
    .groups = "drop"
  ) |> 
  ggplot(aes(gesture, count_tot, colour = background),
         alpha = 0.2) +
  geom_line(aes(group = dyad),
            alpha = 0.5) +
  geom_point() +
  facet_grid(background ~ months) +
  labs(
    title = "Infant gesture counts at 10, 11 and 12 mo",
    x = "Gesture type", y = "Gesture count"
  ) +
  theme_light() +
  theme(legend.position = "none")

Connected dots plot

WALS data

library(ritwals)
wals_data <- intersect_features(c("45A", "70A")) |> 
  filter(feature_ID %in% c("45A", "70A")) |> 
  select(feature_ID, value, language) |> 
  filter(str_detect(value, "No", negate = TRUE)) |> 
  pivot_wider(names_from = feature_ID, values_from = value) |> 
  count(`45A`, `70A`) |> 
  mutate(
    `45A` = factor(`45A`, levels = c("Binary politeness distinction", "Multiple politeness distinctions", "Pronouns avoided for politeness")),
    `70A` = factor(`70A`, levels = c("Second singular and second plural", "Second singular", "Second plural", "Second person number-neutral"))
  ) |> 
  drop_na()
wals_data

# A tibble: 8 × 3
  `45A`                            `70A`                                 n
  <fct>                            <fct>                             <int>
1 Binary politeness distinction    Second person number-neutral         10
2 Binary politeness distinction    Second plural                         1
3 Binary politeness distinction    Second singular                       5
4 Binary politeness distinction    Second singular and second plural    15
5 Multiple politeness distinctions Second person number-neutral          1
6 Multiple politeness distinctions Second singular                       2
7 Multiple politeness distinctions Second singular and second plural     7
8 Pronouns avoided for politeness  Second person number-neutral          2

Alluvial plot

wals_data |> 
  ggplot(aes(y = n, axis1 = `45A`, axis2 = `70A`)) +
  geom_alluvium(aes(fill = `45A`)) +
  geom_stratum() +
  geom_text(stat = "stratum",
            aes(label = after_stat(stratum))) +
  scale_fill_brewer(palette = "Dark2") +
  labs(title = "Languages with a politeness distinction and a morphological imperative") +
  theme_light() +
  theme(legend.position = "none")

Alluvial plot

Violin plot

polite |> 
  drop_na(f0mn) |> 
  group_by(subject) |> 
  mutate(f0mn_z = (f0mn - mean(f0mn)) / sd(f0mn)) |> 
  ggplot(aes(attitude, f0mn_z)) +
  geom_violin() +
  geom_jitter(aes(colour = attitude),
              width = 0.05,
              alpha = 0.5) +
  stat_summary(fun = "median", geom = "point",
               shape = 17, size = 5) +
  facet_grid(cols = vars(musicstudent), labeller = label_both) +
  scale_color_manual(values = c("orange", "purple")) +
  labs(
    title = "Mean f0 of Korean speakers by attitude",
    y = "Mean f0 (z-scores)",
    caption = "Korean speakers were students living in Germany. Mean f0 was z-scored within each speaker."
  ) +
  ThemePark::theme_starwars() +
  theme(legend.position = "none")

Violin plot

Don’t use box-plots

The Datasaurus example.

Also see post by Nick Desbarats: I’ve stopped using box plots, should you?.

Don’t use error bars

See The issue with error bars by Data To Viz.

Don’t plot means of proportion/percentage means

set.seed(992)
participants <- 10
observations <- round(runif(participants, 1, 15))
probabilities <- runif(participants)
binary_outcomes <- rbinom(participants, size = observations, prob = probabilities)

mean(binary_outcomes / observations)

[1] 0.714881

sum(binary_outcomes) / sum(observations)

[1] 0.7866667

• The mean of means is the same as the overall mean only if the N for each participant is the same.

• Even in this case, a mean by itself is MEANingless and plotting standard errors with error bars can be misleading (see previous slide).

• So plot the proportion of each participant and the overall proportion!

To summarise

1. Show raw data (e.g. individual observations, participants, items…).

2. Do not use box-plots, even if you are asked to!

   1. Use density plots, violin plots, strip charts.

3. Do not use “error bars”, even if you are asked to!

   1. If the person asking insists, ask “which measure should I use for the error bars?”.

   2. The only sensible answer is “X% Confidence Intervals” (where X is usually 95%).

4. Do not use means of proportions!

   1. Calculate the overall proportion and/or show the proportion of each participant as raw data.