Describing the Data [R]

What we are building today

Working session

• Three figures and one descriptives table, from last week’s analysis table
• A line for change over time, a bar for counts, a histogram for shape
• Each is a different geometry chosen for a different kind of question
• These figures go straight into your White Paper’s Results and onto your poster
• Plus the reporting rule that turns a picture into a claim: the effect size

Thirty-five thousand rows you cannot read

• The analysis table is complete, faithful, and almost impossible to read
• Nobody scrolls 35,267 rows and comes away knowing whether gaming chat runs longer
• The numbers are all present. The pattern is not visible.
• A figure turns a column too long to hold in mind into a shape the eye reads at once
• Visualization is not decoration applied after the analysis. It is part of it.

The grammar of graphics

Three parts do most of the work (Wilkinson, 2005; Wickham, 2016):

ggplot(data, aes(x = ..., y = ...)) +
  geom_line()

• Data: the table being shown
• Aesthetic mappings: which column goes to which visual property
• Geometry: the mark used to draw it, a line, a bar, a point

Swap geom_line() for geom_col() and the same data becomes bars.

Figure 1: a line for change over time

• The question: across the collection week, how did viewership move, and how was it split across games?
• Horizontal axis is time, vertical axis rises and falls, so the geometry is the line chart
• A line implies the points are in sequence and the space between them is real
• That is why it is right for time and wrong for unordered categories
• The data needs shaping first: six-hour buckets, top five categories kept

Shaping the data

viewers_over_time <- streams %>%
  filter(!is.na(game)) %>%
  mutate(
    timestamp = as.POSIXct(date / 1000,
                           origin = "1970-01-01", tz = "UTC"),
    six_hour  = floor_date(timestamp, "6 hours"),
    category  = fct_lump_n(game, n = 5)
  ) %>%
  group_by(six_hour, category) %>%
  summarise(total_viewers = sum(viewers), .groups = "drop")

Bucket the snapshots into six-hour windows, collapse everything past the top five games into “Other”, and sum viewers inside each bucket.

Drawing the line

ggplot(viewers_over_time,
       aes(x = six_hour, y = total_viewers, color = category)) +
  geom_line(linewidth = 0.9) +
  labs(title = "Concurrent viewers by game category",
       x = "Date (UTC, six-hour buckets)",
       y = "Total viewers in bucket",
       color = "Category") +
  v2v::scale_colour_v2v() +
  v2v::theme_v2v()

theme_v2v() applies the fonts, spacing, and gridlines of a publication-ready figure in one call, so all three figures look like one set.

What dominates, and why that is partly your choice

• The line that towers over the plot is “Other”, the catch-all beyond the top five
• One reading is substantive: viewership is genuinely spread across a long tail of games
• The other is a caution about the chart’s own choices: lumping fifty-plus categories into one bucket all but guarantees that bucket is largest
• There is a plain design cost too. The five named lines are pressed into the bottom, so Fortnite against Just Chatting is hard to read.
• A dominant series crowds out the rest. Noticing that is reading a figure honestly.

Figure 2: a bar for counts

• The question: across the hours of the day, when are the channels busiest?
• Hour of day is not a continuous sweep. It is 24 discrete categories.
• What is measured for each is a simple count of messages
• Counts across discrete categories call for the bar chart: one bar per category, height is the count, bars side by side for comparison

The code

chat_by_hour <- analysis %>%
  mutate(hour = hour(timestamp)) %>%
  count(hour, name = "messages")

ggplot(chat_by_hour, aes(x = hour, y = messages)) +
  geom_col(fill = "#2f7d8a") +
  labs(title = "Chat volume by hour of day",
       x = "Hour of day (UTC)", y = "Messages in sample") +
  v2v::theme_v2v()

geom_col() draws a bar whose height is a value already computed, which is exactly what count() produced.

The daily pulse

• Busiest through UTC midday and afternoon, peaking at 13:00 with 2,201 messages
• Quietest in the small hours, bottoming out at 03:00 with 833
• A swing of well over two to one between the loudest hour and the quietest
• The shape is not mysterious: Twitch’s 2018 audience concentrated in the Americas and Europe, and 03:00 UTC is the middle of the night across most of that span
• The point is what the bar chart does: 24 numbers become a rhythm in one glance

Effect sizes, and their standing

“Effect sizes are the most important outcome of empirical studies.”

Lakens (2013, p. 1)

• Every figure that displays a comparison must report the effect size of that comparison
• A bar showing one group higher than another is a visual impression, not a claim
• The effect size quantifies the magnitude in a scale-invariant unit

Figure 3: a histogram for shape

• The study’s central question: how long is a chat message, and does the answer depend on the kind of channel?
• Not a total, not a trend. A distribution: where values cluster and thin.
• The geometry is the histogram: slice the range into equal intervals called bins and draw a bar for how many values fall inside each
• Two histograms overlaid, gaming and non-gaming, so the shapes compare directly

The code

msglen <- analysis %>%
  filter(!is.na(is_gaming)) %>%
  mutate(length_shown = pmin(message_length, 120))

ggplot(msglen, aes(x = length_shown, fill = is_gaming)) +
  geom_histogram(binwidth = 5, position = "identity", alpha = 0.55) +
  labs(x = "Message length (characters, capped at 120)",
       y = "Messages", fill = "Gaming channel") +
  v2v::scale_fill_v2v() +
  v2v::theme_v2v()

Drop the unlabeled messages, cap the displayed length, and overlay the two groups at partial transparency so both shapes stay visible.

Two choices you have to disclose

• binwidth = 5 sets each bar to cover five characters. A wider bin smooths the shape, a narrower one roughens it. The number is a judgment you report.
• pmin(message_length, 120) caps the displayed length, so everything longer lands in the final bar
• Twitch’s real limit is 500 characters, so that last bar is a genuine pile-up
• Capping keeps the bulk legible instead of stretched thin by a few outliers
• A cap that is not announced is a quiet distortion. The axis label says so.

What the shape shows

• Both groups share one shape: heavily right-skewed
• A tall stack of very short messages at the left, a long thin tail to the right
• Twitch chat is mostly brief, a word or an emote, with a minority stretching the range
• This is the most important thing the histogram reveals, and it is invisible in any single summary number

The descriptives

msglen %>%
  group_by(is_gaming) %>%
  summarise(
    n      = n(),
    mean   = mean(message_length, na.rm = TRUE),
    median = median(message_length, na.rm = TRUE),
    sd     = sd(message_length, na.rm = TRUE),
    .groups = "drop"
  ) %>%
  v2v::pretty_table()

Split by gaming status and compute the four numbers that a Results section needs for each group.

The table

• Read the mean column: non-gaming chat averages 33.70 against 28.49
• Read the median column: 16 against 17, all but identical, pointing the other way

When mean and median disagree

• The mean is sensitive to a long tail in a way the median is not
• A small number of very long messages pulls the average up while leaving the middle value untouched
• The non-gaming group has the heavier tail, recorded directly by its standard deviation: 60.68 against 38.47
• The gap in means is real arithmetic, but it is the work of the tail
• Lean on the mean alone and you report that non-gaming chat is longer. The typical message is the same length in both.

What to report, and with which comparison

• Two group means: Cohen’s d with a 95% confidence interval
• A cross-tabulation of two categorical variables: Cramer’s V
• A correlation: r and its confidence interval
• v2v::run_t_test() and v2v::run_chi_square() compute all of these and return output formatted for the caption
• Put the effect size in the figure caption or a companion table, not in a footnote

Figures other people can read

• Roughly one in twelve men has a form of color-vision deficiency
• Choose a colorblind-safe palette, and do not let color carry the message alone: vary line type, or split groups into panels, so the figure survives grayscale
• Alt text is what a screen-reader user gets instead of the figure, and a figure with none is, to them, simply missing
• Good alt text states the chart type, what is on each axis, and the takeaway
• In Quarto it goes in the figure’s fig-alt attribute, and writing it is a test of whether the figure shows anything at all

Checkpoint

You should now have:

• Three rendered figures, each ending in v2v::theme_v2v()
• A descriptives table with n, mean, median, and sd for both groups
• fig-alt text written for all three figures
• The binwidth and the 120-character cap named in your axis labels or captions
• One sentence per figure saying what it shows, ready to paste into Results

Before Week 13

• Submit Describing Data [R] (100 points): figures, descriptives, alt text, and the disclosed choices
• Read Chapter 13 with the graduate toggle on
• Read the assigned article: Lakens et al. (2018), “Justify your alpha”, Nature Human Behaviour, 2, 168 to 171
• Write your journal entry, 450 to 500 words, engaging both
• The histogram left a precise question: is the five-character gap real, or sampling noise? Week 13 answers it.