Plots That Tell the Truth: Visualization Best Practices

A figure is an argument, so it can lie

When you put a chart in a paper or a slide, you are making a claim and asking the viewer to believe it at a glance. That power cuts both ways. Small, "innocent" choices, where the axis starts, whether you show the spread, which colors you pick, can exaggerate a tiny effect into a dramatic one or bury a real signal. The goal of this lesson is a single discipline: let the reader see what the data actually say, and make the honest reading the easy one.

Deception 1: the truncated y-axis

The most common chart crime. A bar chart whose y-axis starts at, say, 98 instead of 0 turns a 2% difference into a bar that looks five times taller than its neighbor. For bar charts, the y-axis must start at zero, because the bar's length is the visual encoding of the value and a truncated baseline lies about that length. (Line charts of a trend can sometimes justify a non-zero baseline, but you must label it clearly.)

Deception 2: the bar chart that hides the data

A bar showing the mean of a group (often with an error bar) throws away almost everything: the sample size, the spread, outliers, and whether the distribution is even unimodal. Two groups can have identical bars and completely different underlying data. The fix is to show the data points: a dot plot, box plot, or violin plot reveals the distribution the bar conceals. With small sample sizes (common in biology), plotting every point is not optional, it is the honest minimum.

See why summaries lie: Anscombe's quartet

In 1973 the statistician Francis Anscombe built four datasets with nearly identical means, variances, and correlations, yet wildly different shapes (a clean line, a curve, a line dragged by one outlier, and a vertical cluster with a single far point). The summary statistics cannot tell them apart; only a plot can. The cell below computes the stats for two of them so you can see the numbers match. Press Run (first run loads Python in your browser).

This is the entire case for showing your data. Identical numbers, different reality. A bar chart of these would be indistinguishable; a scatter plot tells the whole story.

Deception 3: color that misleads

Color is data, so it has to be honest too. Three rules:

• Avoid the rainbow ("jet") colormap. It is not perceptually uniform: equal steps in value look like unequal steps in color, inventing false boundaries and hiding real ones. Use a perceptually uniform map such as viridis for continuous data, the default in modern tools for exactly this reason.
• Design for color-blind readers. Roughly 1 in 12 men cannot distinguish red from green, the single most common chart color pair. Use color-blind-safe palettes (Okabe-Ito, viridis) and reinforce color with shape or labels so the figure still works in grayscale.
• Use diverging palettes for diverging data. A heatmap of log fold changes (negative to positive around zero) should use a two-hue diverging scale centered at zero, not a single-hue ramp that hides the sign.

Deception 4: chartjunk and the wrong chart

Edward Tufte's principle is to maximize the "data-ink ratio": every drop of ink should convey data. 3D bars, heavy gridlines, drop shadows, and decorative backgrounds add nothing and distort comparisons (3D especially makes lengths and angles unreadable). And some chart types are simply weak: pie charts make it hard to compare similar slices (use a bar chart), and dual y-axes let you imply a correlation between two series by choosing the scales, a favorite trick of misleading figures.

The genomics plots, done honestly

Plot	Shows	Do it right
Volcano plot	Effect size (log2FC, x) vs significance (−log10 adjusted p, y)	Use the adjusted p-value; mark both thresholds; don't truncate axes to exaggerate
MA plot	Log fold change vs mean expression	Reveals whether fold changes depend on expression level (a normalization check)
PCA plot	Samples in reduced dimensions	Label the % variance on each axis; color by the real biological group, and check for batch effects
Heatmap	Expression across genes × samples	State whether values are scaled (z-scored) per row; use a diverging, perceptually uniform palette
p-value histogram	Distribution of all raw p-values	Your first diagnostic: flat + spike at 0 is healthy (see the multiple-testing lesson)

Notice these plots are honest by design: a volcano plot forces you to show effect size and significance together, which is exactly the lesson from earlier in this track. Good plot choices bake good statistics into the picture.

A checklist for every figure you make

1. Do bar axes start at zero? If not, fix it or switch chart type.
2. Can the reader see the data, not just a mean? Show points for small n.
3. Is the color scale perceptually uniform and color-blind safe? No rainbow, no red-green-only.
4. Are axes labeled with units, and is any transform (log, z-score) stated?
5. Is there a clear caption with n, the test used, and what the error bars represent?
6. Does the visual emphasis match the real effect size, or is the chart over-dramatizing?

Check your understanding

You finished Track 5

Next up: Track 6, Reproducibility & Career →