Every chart you make is making a point: this went up, this group is worse off, these two things move together. A good chart helps a reader see that point faster than a paragraph could. A bad one hides it, or worse, makes a point the data does not support.

Once you see charts this way, “what chart should I use?” becomes a sharper question: which channel will carry my most important variable?

A bar encodes a value with its length. A reader judges a bar twice as long as carrying twice the value — so the bar must start at zero. Cut the baseline and a 2% difference can look like a doubling. This is the single most common way charts mislead.

Colour is great for showing categories (which line is which) and rough intensity (a heatmap). It is poor for precise quantities — nobody can read “47” off a shade of blue. Colour gets its own section later; for now, two rules:

Suppose you are comparing female literacy across 12 states. The states are the categories; literacy is the number that matters. Put literacy on position or length (a sorted bar or dot plot), and use colour only to flag the one state you want to talk about.

Do not start from “I want a chart.” Start from the question your reader has. Almost every question is one of five kinds, and each kind has a natural chart family.

Time goes on the horizontal axis, left to right. A line connects points to show the trend; the slope is the message. Use a line for continuous time, a bar only when the periods are few and discrete.

A scatter plot puts one variable on each axis and one dot per observation. Clusters, lines and outliers jump out. Add a trend line only if it genuinely helps — and never let it imply causation (more on that soon).

Most charts use a linear scale, where equal distances mean equal amounts. A log scale makes equal distances mean equal multiples (10, 100, 1,000) — useful when values span many orders of magnitude, or when the rate of growth is the story.

You can plot accurate data and still mislead — through the axis, the scale, the time window or the framing. Because charts are read fast and trusted instinctively, a misleading chart does more damage than a misleading sentence. This section is the most important in the course.

The zero rule is about bars, where length carries the value. Lines encode change, so they may start where the data lives — forcing a stock index or a temperature series to include zero can flatten a real, meaningful change.

Two lines, two different y-axes, scaled so they appear to track each other — the reader infers a relationship that the analyst manufactured by choosing the scales. By sliding the axes you can make almost any two series look correlated.

Start the time axis at an unusually low year and the trend looks like a boom; start at a peak and the same series looks like a collapse. The data is real; the window is the lie.

3D bars and pies tilt the geometry so that nearer slices look bigger and the back of the chart shrinks. They add no information and distort the one channel that mattered. The same goes for drop shadows and “glossy” effects.

When a value is shown as a circle or icon, the reader judges it by area. If you set the radius proportional to the value, a figure twice as large draws a circle four times as big. Pictographs (“one person = 1 million”) that scale a single stretched icon make the same error.

A scatter with a trend line, or two lines rising together, strongly suggests that one causes the other. The chart cannot show causation — only association. Ice-cream sales and drownings both rise in summer; neither causes the other.

For a quantity that runs low to high — poverty rate, temperature, density — use a single hue from pale to deep. Darker reads as “more” intuitively. Keep the steps evenly spaced in perceived lightness so equal jumps look equal.

When values spread both ways from a natural centre — above/below average, gain/loss, agree/disagree — a diverging scale puts a neutral colour at the midpoint and two contrasting hues at the ends. The midpoint must be the real zero or mean, not an arbitrary value.

For categories with no order, pick hues that are easy to tell apart and roughly equal in weight (so none shouts). Past about seven colours, readers lose track and the legend becomes a memory test.

Roughly 8% of men and 0.5% of women have some colour-vision deficiency, most commonly red–green. A chart that relies on red-vs-green to make its point fails for millions of readers.

The classic “jet” rainbow ramp (blue–green–yellow–red) looks lively but lies: it has bright bands that invent boundaries where the data is smooth, and dark ones that hide real differences. It is also poor for colour-blind readers.

Marks show the pattern; words tell the reader what it means and what to trust. Titles, labels, annotations, units and sources do as much work as the geometry. A beautiful chart with no words is a puzzle.

A separate legend forces the reader’s eye to bounce between the key and the chart, matching colours from memory. Where you can, write the label next to the thing it names — at the end of each line, beside each bar.

An annotation — a short note with an arrow or a marked point — turns a chart from “here is some data” into “here is what happened.” Mark the spike, name the policy, flag the outlier. It is the cheapest way to make a chart memorable.

“63.7%” implies you know the figure to a tenth of a percent; often you do not. Round to the precision your data and your reader actually need — usually whole numbers for a public audience.

When readers need exact figures, when there are only a handful of numbers, or when they will look up specific rows, a well-set table beats a chart. Tufte’s rule of thumb: for a small dataset that will be read closely, a table often communicates better.

A sparkline is a small, axis-free line drawn inline — enough to show a trend right next to the number it describes. Add a column of sparklines to a table and readers get the exact value and the shape of its history together.

Sometimes the most effective “visualisation” is a single large figure with a label and a tiny bit of context. Used on dashboards and report covers, these big-number cards make the headline impossible to miss.

A histogram groups values into bins and shows how many fall in each. It reveals shape — symmetric, skewed, bimodal. But the bin width changes the story: too wide hides structure, too narrow turns signal into noise.

A box plot compresses a distribution into median, quartiles and whiskers — great for comparing many groups at once. The cost: it hides the actual shape. Two very different distributions can produce identical boxes.

When you have room and not too many observations, plot every point. A strip plot lines them up; jittering spreads overlaps apart; a beeswarm packs them into a shape that doubles as a density. Readers see the spread, the clusters and the outliers — nothing is hidden behind a summary.

Most development figures are estimates from samples or models. A point drawn without its uncertainty looks more certain than it is. Error bars, confidence bands and ranges tell the reader how much to trust the dot.

A forecast drawn as a single confident line invites the reader to believe a precision that does not exist. Show projections as a fan — a widening band — so the growing uncertainty is visible. Mark clearly where measured data ends and the projection begins.

A map feels authoritative and is genuinely useful when location is the story. But a lot of data plotted on a map would be clearer as a bar chart — the geography adds beauty while making values harder to compare.

A choropleth shades each region by a value. Its biggest trap: shading by a count rather than a rate. A map of “total cases” mostly shows where the people are — big, populous states light up regardless of how bad things actually are there.

Normalising turns a map of population into a map of the thing you care about. Deaths → deaths per 100,000. Spending → spending per person. Schools → schools per 1,000 children. The denominator is a design decision — choose the one that matches the question.

On a normal map, area equals visual weight. A huge, sparsely populated state (Rajasthan, Ladakh) dominates the eye, while a tiny, densely populated one (Delhi, Kerala’s coast) all but disappears — even when far more people live there. The map over-weights land and under-weights people.

A cartogram distorts regions so their size reflects a value (population, electorate) instead of land area. A hex cartogram gives every region an equal tile — useful when you want each state read equally, regardless of size.

A choropleth groups values into colour bins, and the cut-points are a choice. Equal-interval bins split the range evenly; quantile bins put equal numbers of regions in each colour. The same data can look alarming or calm depending on the scheme.

A chart for a statistics-literate panel can carry more complexity than one for the general public. A chart for a printed report is read slowly and closely; one on a projector has three seconds from the back row. Design for the actual reader and the actual medium.

Because some readers cannot distinguish certain colours — and because charts get printed in greyscale and photographed badly — pair every colour with a second cue. Then the chart still works when the colour is lost.

Chartjunk is decoration that adds no information and often subtracts clarity: 3D effects, gradients, clip-art, busy backgrounds, needless gridlines, redundant legends. It makes a chart look “designed” while making it harder to read.

Interaction (hover, filter, zoom) and animation (transitions between states) are powerful for exploration and for showing change. But a static reader sees none of it, and a gratuitous animation just delays the point. Always make the chart work as a still image first.

Before opening any tool, sketch the chart on paper. It costs seconds, forces you to decide what goes on each axis, and surfaces a better idea than the software’s default. The tool should execute your decision, not make it for you.

Theory only takes you so far. Pick one chart a week — from a newspaper, a report, your own work — and redraw it better. Decide what it is for, fix the encoding, cut the junk, write a finding title, add the source. Keep the before-and-after.