Notes from reviewing a lot of dashboards. This one is about a single setting that causes more silent damage than any formula: where the vertical axis begins.

A bar is a length

When people look at bars, they compare lengths. Not labels, not numbers. Lengths. So the length has to mean the value, and that only works if the bar starts at zero.

Here is a worked case. Customer satisfaction last quarter was 100 (index) and this quarter it is 96. A fall of 4%. Now draw two bars with the axis starting at 90. The old bar is 10 units tall and the new one is 6. The new bar looks 40% shorter. A reader who glances at that chart will think the score has collapsed. I checked the arithmetic: 6 divided by 10 is 0.6, so the visual drop is ten times the real one.

Spreadsheet and BI tools often choose “automatic” axis bounds, which hug the data. If your values are 96 and 100, the software helpfully zooms in. That is fine for finding detail. It is a bad default for a chart that will be shown to a director in a three-second glance.

My rule, stated as opinion: any bar chart that goes outside the team starts at zero, with no exceptions I would defend in a review.

When a non-zero start is fine

The rule is about encoding, not about zero being sacred. Different marks encode differently.

  • Line charts encode by position and slope. A stock-price or daily-sales line that runs between 96 and 104 can use an axis from 90 to 110, because the eye is reading shape and change, not length. Starting at zero would flatten a real pattern into a ruler. Label the axis clearly, though, so nobody reads it as a crash.
  • Dot plots and range charts, where a dot marks a value, do not need zero either. Many analysts prefer them for comparing close values for exactly this reason.
  • Measures with a natural non-zero scale, such as body temperature or exchange rates, make zero meaningless anyway.

If the real question is “are these two values close?”, a bar chart is the wrong tool in the first place. Use a table with the difference computed, or a dot plot, and let the reader see the gap.

The other axis tricks

Truncation is the famous one. These deserve equal suspicion.

Dual axes. Put revenue (in ₹ crore) on the left and orders (in thousands) on the right and you can make them cross wherever you like by rescaling one. Any apparent crossover or “correlation” is an artefact of the scales you picked. If two series need two axes to fit on one chart, consider two small charts stacked and sharing a time axis. They are clearer, and nobody can accuse you of choreography.

Unequal intervals. A time axis with monthly points up to December and then quarterly points after that bends the slope. Keep intervals even or show the gaps.

Log scales. A log axis is legitimate for growth rates and for data spanning several orders of magnitude, since equal distances mean equal percentage changes. But many audiences read it as linear. If you use one, say so in the title, not in a footnote.

Cropped or clipped bars. One outlier branch with ₹9 crore of sales dwarfs the others at ₹40 lakh to ₹80 lakh. Cutting that one bar with a zigzag is sometimes acceptable if clearly marked, but it is usually better to show the real bar and let the dull ones be dull, or to give the outlier a separate annotation.

Questions to ask before sending a chart

  1. What mark am I using, and does it encode by length or by position?
  2. If I were hostile, could I get a different impression by changing only the axis? If yes, which version is the honest one?
  3. Is the difference I want to show large enough to show on a zero-based axis? If it is not visible there, maybe it is not big enough to matter. State it in a number instead.
  4. Are the data labels on? Putting exact values on the bars removes a lot of room for visual argument.

Number three is uncomfortable, and I like it for that. Small differences that vanish at honest scale are often noise anyway. If a 4% movement matters to the business, put “down 4%” in the title and use the chart for context.

A last point on intent. Most truncated bars in internal reports are not lies; they are defaults nobody looked at. But a chart’s effect does not depend on the author’s intent. Treat the axis as part of the message, because it is.

Figures and names in the examples are illustrative.

Leave a Reply

Trending