learn-dataviz-with-phoebe / Session 2 of 6
Learn Data Visualization with Phoebe · Session 2 of 6

Choosing the form

Most people pick a chart type the way they pick a font - from what is at the top of the menu and what they used last time. Tonight the choice comes from the job you named in session 1, which narrows twenty options to about two. Plus the cheapest improvement in this entire discipline, which is sorting, and the one rule here that has no exceptions at all: two measures never share a chart with two y-scales.

🟢 Easy Anyone who defaults to whatever the tool opened with Bring a chart you are unsure about 45 min
0-3 · Welcome 3-18 · The shortlist, sorting, and the dual-axis rule 18-40 · Build-along: re-form four charts 40-45 · Q&A
Part 0

Where we are

Session 1 gave every chart a written question and a named job. Tonight the job picks the shape, and it turns out to pick it fairly narrowly.

🎫 Clear all three quiz questions to stamp your passport Live = presented in the 45 minutes Self-study = the full depth, after class

★ What you walk out with tonight. Four Daybreak charts re-formed from their jobs, one dual axis killed, and a sorted bar chart that is better than its unsorted version for free.

Part 1 · covers the shortlist

The job picks the shape

6 min live
Form Good at Bad at Bar comparing magnitudes precisely many categories, or time with gaps Line change and rate of change unordered categories - implies a sequence Dot many categories, sparse values part-to-whole, and small differences Heatmap a matrix - cohort against month reading exact values off a colour Stacked bar the total, and the biggest part comparing the middle segments Pie two or three parts of one whole four or more parts. Almost everything else Only the last column is ever surprising. Every form is bad at something, and the way to choose is to check that the thing it is bad at is not the thing your question was about.
🔍 Click to zoom - pick by what it is bad at, not by what it is good at
LiveSorting, which is free and almost never done4 min▶

A categorical bar chart in the order the data happened to come out of the database asks the reader to scan every bar and hold the largest so far in their head. Sorted by value, the answer is the top row.

  • Sort by value for magnitude questions. Which is biggest is now a glance rather than a scan.
  • Keep the natural order when there is one. Months, sizes, severity levels, survey scales. Sorting those by value destroys the thing the reader uses to orient.
  • Keep the order stable across a set. If four charts share categories, sort them once, by the same key, or the reader re-learns the layout four times.

This is the highest ratio of improvement to effort anywhere in the course. It changes no data, costs no space, needs no colour decision, and it is missing from most charts in most organisations because the default is whatever order the query returned.

Part 2 · covers the one hard rule

Never two y-axes

6 min live

This is the only rule in this course stated without exceptions, and the reason is not aesthetic. A dual-axis chart lets the author produce whatever relationship they want between two series, by choosing two scales, and the reader has no way to see it happening.

Same data. Right-hand scale 0 to 500. Reads as: the two move in opposite directions. Same data. Right-hand scale 0 to 5,000. Reads as: one is flat while the other climbs. Neither chart contains a false number. The author chose a second scale, and the choice produced the relationship. A reader cannot audit it, because there is no correct second scale to compare it against - which is exactly why this is the one rule with no exceptions rather than a preference.
🔍 Click to zoom - the same two series, two defensible scales, two opposite conclusions
LiveWhat to do instead, in three moves4 min▶
  • Two charts, stacked, sharing an x-axis. The simplest fix and almost always the right one. The reader can still see the timing relationship, and neither series is being scaled to flatter the other.
  • Index both to a common base. Set each series to 100 at the start and plot the change. Now one axis is honest for both, and the question "which grew faster" is answerable directly.
  • Plot the thing you actually mean. Often the dual axis exists because the real question is a ratio - revenue per subscriber, cost per order. Compute it and plot one series.
The third one is the interesting one

When somebody puts signups and revenue on two axes, the question in their head is usually "is each new subscriber worth as much as the last one?" That is revenue per subscriber - a single series, one axis, and it answers the question the dual axis was gesturing at without letting anybody choose the answer.

Self-studySmall multiples, and when eight series becomes eight charts4 min▶

Past about four series, a line chart becomes a plate of spaghetti and the legend becomes a matching exercise. The fix is not a better palette - it is more charts.

  • Small multiples: the same chart repeated once per category, on shared axes, small. The eye compares shapes across the grid far better than it untangles overlapping lines.
  • Shared scales are what make it work. If each panel gets its own y-axis you have made eight unrelated charts that look comparable, which is worse than the spaghetti.
  • Grey the rest. When one series is the point, draw it in colour and the other seven in light grey on the same axes. Context without competition, and it needs no legend.

The grey-the-rest move is worth having as a reflex. It answers "but I need to show all of them" without giving eight series equal weight, and it is one line of code in any library.

Build-along

Re-form four Daybreak charts

★ 22 min · everyone builds

Four questions from session 1, four forms to choose. Sketch them on paper - the point is the decision, not the rendering.

Job. From session 1's five. Magnitude, change, identity, polarity, one number.

Shortlist. The job leaves you two candidates, not twenty. Magnitude gives bar or dot; change gives line.

Bad at. Check what each candidate is bad at, and whether that is your question. A stacked bar is bad at middle segments - if your question is about a middle segment, it is out.

Sort. By value for magnitude; by the natural order where one exists; the same way across a set.

Check the axes. One y-axis. If you find yourself wanting two, you have found either two charts or a ratio you have not computed yet.

Chart one: "Which roast earns most in a month?" Magnitude, four categories. Sorted bar, one series, one colour, axis from zero. This is the bench's chart in session 3.

Chart two: "Is the subscriber base growing?" Change over time. Line. Note that it does not have to start at zero - and note that you now need session 3 to explain why that is fine here and a distortion on the bar chart.

Chart three: "Which cohorts churn fastest?" A matrix - cohort by month. Heatmap or small multiples. Decide which by asking whether the reader needs exact values, because a heatmap cannot give them.

Chart four: "Are new subscribers worth as much as old ones?" This is the trap. It looks like two series - signups and revenue - and it is one: revenue per subscriber. Compute it and plot one line.

Now find a real dual axis. In your own work or something you were sent this week. They are everywhere. Decide which of the three fixes applies.

Sort one real chart and look at the before and after side by side. This is the moment most people become permanently annoyed by unsorted bar charts.

Take one spaghetti line chart and grey the rest. One series in colour, the others in light grey, no legend. Then ask whether you have lost anything.

If you take one thing from tonight. When you want a second y-axis, the question you actually have is usually a ratio. Compute it and plot one series - it answers the real question and it removes the reader's inability to audit you.

Homework

Before session 3

25 min
  1. Kill one dual axis. Find a real one, pick the fix, and rebuild it. Note which fix you needed.
  2. Sort three bar charts and look at each before and after.
  3. Convert one spaghetti chart to small multiples or grey-the-rest.
  4. Find a chart whose form fights its question. A line over unordered categories, a stacked bar where the middle segment is the point, a pie with seven slices.
  5. Bring your sorted bar chart to session 3, because session 3 is going to ask whether its axis starts at zero and then measure what happens when it does not.
Check yourself

Three questions

Clear all three and session 2 gets stamped in your passport.

1 · Why is "never two y-axes" stated without exceptions rather than as a preference?

The same two series on two defensible right-hand scales can be made to look inversely related or unrelated, with no false number anywhere in either chart. That is not clutter, it is an unauditable claim - which is why it gets a rule rather than a guideline.

2 · Somebody wants signups and revenue on one chart with two axes. What is the question they probably have?

Dual axes usually appear where the real question is a ratio nobody has computed. Working out what that ratio is answers the question directly, removes the scale choice, and produces a chart the reader can actually check. B is a real question too, and its fix is two stacked charts sharing an x-axis.

3 · When should you NOT sort a bar chart by value?

Sorting is the cheapest improvement available and it has exactly one common exception: an inherent order. Sorting months by value makes a chart nobody can read. Appearing alongside other charts is an argument for sorting them all the same way, not for leaving them unsorted.

Sources covered

What this session teaches, and where it came from

✓Form by jobBar, line, dot, heatmap, stacked bar and pie, each with what it is good at and - the useful column - what it is bad at
✓Never a dual axisShown as two defensible scales producing two opposite conclusions from identical data, with the three replacements
✓SortingBy value for magnitude, natural order where one exists, the same key across a set. Highest improvement-to-effort ratio in the course
✓Small multiples and grey-the-restPast four series the fix is more charts, not a better palette - with shared scales as the thing that makes it work
✓Pie charts, honestlyTwo or three parts of one whole. A bar is better for almost everything else
◐Chart libraries and toolingThe method is tool-agnostic, and every library can draw all of these badly

Every fact on this page, and its verification tier, is recorded in the course's source map.

Session 2 cheat sheet · pin this

Choose by what it is bad atEvery form fails at something. Check that it is not your question.
Never two y-axesThe scale choice makes the relationship, and nobody can audit it.
The dual axis is a ratioRevenue per subscriber. Compute it, plot one series.
Sort itFree, changes no data, missing from most charts everywhere.
Except natural ordersMonths, sizes, severity, scales. Sorting those destroys orientation.
Grey the restOne series in colour, seven in grey. Context without competition, no legend.