learn-dataviz-with-phoebe / Session 3 of 6
Learn Data Visualization with Phoebe · Session 3 of 6

Honesty

Almost nobody sets out to mislead with a chart. What happens instead is that somebody asks for it to be clearer or bolder, a default gets changed, and the picture quietly stops being proportional to the numbers. Tonight that stops being a matter of opinion: a bench draws a real chart from real data and computes how much the picture is overstating, from the rectangles it just drew. Daybreak's roasts differ by 21 percent. One ordinary change shows it as 173.

🟡 Medium The bench session - everything runs in your browser Bring your sorted bar chart from session 2 45 min
0-3 · Welcome 3-18 · The lie factor, and the four ordinary distortions 18-40 · Build-along: the bench 40-45 · Q&A
Part 0

Where we are

You have a question, a job and a form. Tonight is the first time the course can tell you that you are wrong with a number rather than an opinion.

🎫 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. A lie factor for a chart you own, the four distortions that produce it, and a sentence for the next time somebody asks you to make one pop.

LiveThe lie factor, and what it does and does not catch3 min▶

Tufte's definition, and it is simple enough to do in your head:

lie factor = size of the effect shown in the graphic ÷ size of the effect in the data

An honest chart scores 1.00. Above 1 it is overstating; below 1 it is understating, which is also a distortion and gets noticed far less often.

What it does not catch, said plainly because a number that seems to certify honesty is dangerous:

  • It compares the largest and smallest item. A chart can score exactly 1.00 and still mislead by which window it plots, which denominator it uses, or which comparison it invites.
  • It says nothing about whether the data is right, whether the categories are fair, or whether the question was worth asking.

So it is one check, not a certificate. It happens to be the check that catches the single most common distortion in professional charts, which is why it is worth computing.

Part 1 · covers the four

Four ordinary things that move the number

6 min live
The change Why somebody made it Lie factor Nothing - the honest chart it is what the data says 1.00 Axis starts at 30,000 "to show the difference clearly" 8.32 Scale by area, not length it looked more designed 0.70 3D perspective the template had it no change, and unreadable Gradient fill it looked flat otherwise no change, and costs contrast Row three is the one nobody talks about. Area encoding UNDERSTATES, and an under-stating chart is never accused of lying.
🔍 Click to zoom - every one of these was somebody trying to make the chart better
LiveWhen does an axis have to start at zero?5 min▶

This is the question people argue about, and it has a clean answer that depends on the encoding rather than on taste.

  • Bars: yes, always. A bar encodes value as length. Length is only meaningful from zero, so a truncated bar chart is making a claim about proportion that is not true. This is not a style preference and the bench measures the consequence.
  • Lines: no, and usually should not. A line encodes change. Its job is the slope, not the distance from an origin. Forcing a temperature series or a retention curve to start at zero flattens the thing the reader came for.
  • Dots: no. Same reasoning as lines - position is being read against neighbours, not against an origin.

So "always start at zero" is wrong as a blanket rule and right for bars specifically, which is where nearly all the damage happens. If somebody quotes the blanket version at you about a line chart, the useful reply is: "what is the mark encoding - length or position?"

Self-studyThe distortions the lie factor cannot see4 min▶

All of these produce a chart scoring exactly 1.00, and all of them are more common than a truncated axis in work that has been reviewed:

  • The window. The same series over 3 months, 12 months and 5 years tells three different stories, all true. Choosing the one that suits you is the most common honest-looking distortion there is. Show the longest window you have, or say why you did not.
  • The denominator. Absolute cancellations rose. Cancellations per thousand subscribers fell. Both are facts, and which one you plot is an argument, not a formatting decision.
  • The missing baseline. "Revenue up 21 percent" against what - last month, last year, forecast, the rest of the market? A chart with no comparison invites the reader to supply a flattering one.
  • The excluded category. Four roasts shown, six exist. Nothing in the chart says so.

The habit that catches all four: write, in one line under the chart, what you chose and why. "Twelve months, per thousand subscribers, all six roasts." It takes ten seconds and it is the difference between a chart and a claim.

Build-along

The lie-factor bench

★ 22 min · everyone builds

One real dataset - Daybreak's revenue by roast for a month - drawn as a real chart. Every number under it is computed from the chart as it currently stands: the lie factor from the geometry, the contrast from the rendered colours, the colour separation through a real dichromacy transform.

Lie factor 1.00 · contrast 4.42:1 · palette ΔE 22.9. The honest chart: sorted, direct labels, receded grid, axis from zero. The data spread is 21 percent and the picture shows 21 percent. This is the only state in which a reader can trust their eyes.

Lie factor 8.32. Start the axis at 30,000. The data still says 21 percent; the picture now says 173 percent. Nothing was falsified - a default was changed, for a reason that sounded like clarity.

Lie factor 0.70. Scale by area instead of length. Now the chart understates: 21 percent shown as 15. Distortion runs both ways, and nobody has ever complained that a chart made a difference look too small.

Make it pop: lie factor 8.32. One click switches on three things - truncated axis, 3D, gradient - because that is what the request means in practice. It looks more impressive and it is eight times overstated.

Vision modes. Switch the palette through deuteranopia and protanopia. Two hues that sit 22.9 apart for most readers sit 10.0 and 9.5 apart there. Still above the floor of 8, and visibly compressed - which is session 4.

Press "the honest chart" and read the three meters. 1.00, 4.42:1, 22.9. Then look at the chart - it is plain, sorted, labelled, and slightly boring. That is what honest looks like and it is worth sitting with.

Strip it back and switch on the truncated axis alone. Watch the lie factor go to 8.32 and read the sentence underneath: 21 percent in the data, 173 in the picture. This is the single most common distortion in professional charts.

Now the area lever, alone. 0.70. The chart is now under-selling the difference. Ask yourself when you last heard anybody object to a chart on those grounds.

Build the honest chart, then switch on "make it pop". Notice it turns on three levers at once, because that is what the request means. The score goes from 1.00 to 8.32 and the chart looks considerably more impressive.

Run the four vision modes on the palette strip. The colours visibly converge under deuteranopia and protanopia, and the measured separation drops from 22.9 to about 10. Session 4 is entirely about that number.

Open the table view. It is under the chart, it is four rows, and it is what discharges a contrast warning. A chart with a table view has no accessibility argument left to lose.

Now compute the lie factor on your own chart, by hand. Measure the longest bar and the shortest with a ruler or a screenshot. Divide the graphic ratio by the data ratio. It takes two minutes and it is the most useful two minutes in this session.

Write the one-line provenance under one real chart. "Twelve months, per thousand subscribers, all six roasts." Ten seconds, and it closes the four distortions the lie factor cannot see.

The sentence the bench gives you. Not "that would be misleading", which is an accusation and starts an argument. Instead: "That version shows a 21 percent difference as 173. Happy to make it bolder - can we do it with colour and a direct label instead?" A measurement, then an offer.

Homework

Before session 4

25 min
  1. Measure the lie factor on three real charts - yours, or ones you were sent. A ruler on a screenshot is a perfectly good instrument.
  2. Find a truncated bar chart in the wild. News, a vendor deck, a product dashboard. They take about four minutes to find.
  3. Find an under-stating chart. Harder, and worth the effort - area encodings, bubble charts, anything sized by radius.
  4. Add the one-line provenance to a chart you are shipping this week. Window, denominator, what is excluded.
  5. Bring a palette to session 4 - your team's, your product's, or the one your tool defaults to. Session 4 measures it.
Check yourself

Three questions

Clear all three and session 3 gets stamped in your passport.

1 · Must a chart's axis always start at zero?

The answer follows from the mark, not from taste or from the data. Forcing a retention curve to start at zero flattens the slope the reader came for. Truncating a bar chart breaks the proportionality a bar is claiming. The useful question when somebody quotes the blanket rule is "what is the mark encoding - length or position?"

2 · The bench reports a lie factor of 0.70 for the area-encoded chart. What does that mean?

Below 1.00 means the picture is smaller than the fact. It is a genuine distortion and it is almost never challenged, because nobody complains that a chart made a difference look too modest. Scaling both dimensions by the value is the usual cause.

3 · A chart scores exactly 1.00. What has that NOT told you?

The lie factor compares the largest and smallest item, so it catches the geometry and nothing else. Plotting three months instead of five years, absolute counts instead of a rate, or four of six categories all score 1.00. The one-line provenance under the chart - window, denominator, exclusions - is what closes that gap.

Sources covered

What this session teaches, and where it came from

✓The lie factor (Tufte)Effect in the graphic over effect in the data, computed live from the drawn geometry - plus an explicit statement of what it cannot catch
✓Truncated axesMeasured at 8.32: a 21 percent spread shown as 173 percent, with no false number anywhere in the chart
✓Zero-baseline, resolved by the markBars always, lines and dots not - because one encodes length and the others encode position
✓Area encoding understatesMeasured at 0.70. Distortion in the direction nobody complains about
✓The four distortions a 1.00 chart still allowsWindow, denominator, missing baseline, excluded category - and the one-line provenance that closes them
✓The anti-lever"Make it pop" as three levers at once, and the measurement-then-offer sentence it gives you
◐Deliberate deceptionThis session is about the distortions honest people ship by accident, which are far more common and much harder to see

Every constant, the palette validation report, and the honest limits of each measurement: the course's source map.

Session 3 cheat sheet · pin this

Lie factorEffect shown ÷ effect in the data. Honest is 1.00.
21 becomes 173One truncated axis, no false numbers, lie factor 8.32.
Zero depends on the markBars yes, always. Lines and dots no. Length against position.
Understating is distortion tooArea encoding, 0.70, and nobody ever objects.
1.00 is not a certificateWindow, denominator, baseline, exclusions all survive it.
The sentence"That shows 21 percent as 173. Can we be bolder with colour instead?"