DataslopePrinciples of Visualization
Before you learn a plotting library, learn what makes a chart good. A short tour of the timeless rules: encode well, declutter ruthlessly, tell one story.
A good chart is a thought made visible. A bad chart is a distraction at best, a lie at worst. Both use the same software, the same data, even sometimes the same chart type. The difference is in the choices — and most of the choices follow a handful of principles that have been refined over a century.
This page is about those principles. We won't write much code on this page. The next page (ggplot2) gives you the tools to apply what you learn here.
Charts are encodings
Every chart maps a variable in the data to a visual property on the page. The toolbox of visual properties is small:
| Visual property | Best for | Notes |
|---|---|---|
| Position (x, y) | numeric variables, comparison | by far the most accurate channel |
| Length | numeric, compared to a common baseline | bars |
| Angle / area | numeric (poorly) | avoid pie charts when accuracy matters |
| Color (hue) | categorical | ~7 distinct values is a hard limit |
| Color (intensity) | numeric (ordered) | sequential scales |
| Shape | categorical | low precision; limit to a few values |
| Size | numeric | low precision; only for emphasis |
The most important rule: encode your most important variable in position. People read position with near-perfect accuracy and color/size only crudely.
Match the chart type to the question
Different questions call for different charts:
| Question | Best chart |
|---|---|
| How is one numeric variable distributed? | histogram, density |
| How do two numeric variables relate? | scatterplot |
| How does a numeric variable differ by group? | boxplot, violin, dot plot |
| How do counts compare across categories? | bar chart |
| How does something change over time? | line chart |
| How is a whole made of parts? | bar chart (yes — almost always better than a pie) |
Bar charts are dramatically underrated. Pie charts and 3D charts are wildly overrated. The eye is bad at angles and worse at volumes.
Less is more: the data-ink ratio
Edward Tufte coined the term data-ink ratio: the proportion of a chart's ink that is actually showing data. Maximize it. Ruthlessly remove:
- 3D effects on 2D data
- Background colors that don't encode anything
- Drop shadows
- Heavy gridlines (light is fine; thick is noise)
- Decorative icons
- Redundant labels and legends
A chart's job is to show data. Everything else is competition.
Use color with intention
Color is powerful and easily abused.
- Categorical hues (red, blue, green, ...): use to distinguish unordered categories. Keep the count low — most people can hold ~7 colors apart, and many fewer if the chart is small or the dots are tiny.
- Sequential color scales (light → dark single hue): use for ordered numeric variables. Light = low, dark = high.
- Diverging scales (blue → white → red, etc.): use when the variable has a meaningful midpoint (zero, average).
Two additional rules:
- Color is not free. Each color you add is one more thing the reader has to translate. If you can encode something structurally instead (with position, faceting, or order), prefer that.
- Be colorblind-friendly. Avoid red/green as the only
distinguishing pair. Use palettes like
viridisorRColorBrewer's color-safe sets — modern R plotting tools make this almost free.
Order matters
If your categories don't have a natural order (alphabetic doesn't count), sort them by the thing you're showing. A bar chart of sales by region, sorted by sales, is dramatically clearer than the same chart sorted alphabetically. The eye can read "decreasing in this direction" without effort.
Aspect ratio and scale
- Don't truncate axes for bar charts. (Truncating amplifies small differences and is a classic chart crime.)
- Don't add a meaningless baseline to line charts. Starting a y-axis at zero is the default for bars but a choice for lines.
- Use the right aspect ratio so trends look the way they actually are. Stretching a chart wide flattens trends; squishing it tall exaggerates them.
Title, axis labels, units
This sounds obvious. It is, and yet 80% of beginner charts skip at least one:
- Title that says what the chart is showing, not just the variable names ("MPG falls as weight rises" is better than "mpg vs wt")
- Axis labels with units ("Weight (1000s of lbs)" not just "wt")
- Legend that doesn't restate something already obvious
If a chart needs paragraphs of caption to make sense, the chart isn't doing its job.
One chart, one message
Cram more than one main message into a chart, and most readers will absorb none of them. If you have two things to say, make two charts. The eye is a serial reader; charts are best when they are crisp and singular.
A bad chart and a good chart, same data — look at them and pick the one you'd put in a report.
What changed:
- Sorted (so the order encodes ranking)
- Horizontal (so the names fit and read naturally)
- Single color (no false categorical distinctions)
- Title that says what the chart is about
- Axis label with the right name
The data is the same. The reader's experience is not.
Test your understanding
Which visual property is read with the highest accuracy by the human eye?
Color hue
Position along a common scale
Area
Shape
Why are pie charts often a poor choice?
They take more pixels.
They cannot show more than two categories.
The eye reads angles and areas poorly — bar charts (which use position/length) are usually clearer and more accurate.
R does not support them.
If you have a bar chart with five regions and no inherent order, you should:
Sort alphabetically.
Sort by region ID.
Sort by the value you're plotting, so the chart's order itself encodes a ranking.
Leave them in a random order.
The principles are language-agnostic. The next page is about
the language that makes them easy to apply in R: the grammar
of graphics as implemented in ggplot2.
Relationships Between Variables
Scatterplots, correlations, and cross-tabulations — the toolkit for asking "does X have anything to do with Y?" and (importantly) interpreting the answer carefully.
The ggplot2 Grammar
ggplot2's central idea — that every plot is data + aesthetic mapping + geometry + scales + faceting — and how that idea makes building rich, principled visualizations almost mechanical.