7. Bar charts — the commonest chart, and the commonest mistake · Dr. Seonggu Sohn's Data Visualization, First Steps

The most-used chart in the paper: 14 of the 62 panels. And every time you draw one there are five things to decide.

Decision	Default answer
Order	by value. Alphabetical order carries no meaning
Direction	long names → horizontal bars
Stacking	only when the total means something. Otherwise put them side by side
Baseline	zero. No exceptions
Error bars	say what they are — SD? SEM? 95% CI?

TIL infiltration across thirteen tumour types. Change the order, direction, stacking and baseline.

Loading…

On the right you see where LUAD sits at once. Same data, same kind of chart — only the order changed.

Python Runs in your browser

import matplotlib.pyplot as plt
import viz, numpy as np, json

F = json.load(open("viz_facts.json"))
p = F["pan"]
names = [d["t"] for d in p]
vals  = [d["mean"] for d in p]

fig, ax = plt.subplots(1, 2, figsize=(10.5, 3.6))

# alphabetical
o = np.argsort(names)
ax[0].bar([names[i] for i in o], [vals[i] for i in o], color=viz.GREY)
ax[0].set_title("sorted by name", fontsize=9)

# by value
o = np.argsort(vals)[::-1]
cols = [viz.PAL[0] if names[i] != "LUAD" else viz.PAL[1] for i in o]
ax[1].bar([names[i] for i in o], [vals[i] for i in o], color=cols)
ax[1].set_title("sorted by value (LUAD highlighted)", fontsize=9)

for a in ax:
    a.tick_params(axis="x", rotation=90, labelsize=7)
    a.set_ylabel("mean TIL %", fontsize=8); viz.bare(a)
fig.tight_layout()
print("LUAD ranks %d of the 13 tumour types." % (sorted(vals, reverse=True).index(
      [d["mean"] for d in p if d["t"] == "LUAD"][0]) + 1))

Stacked bars need particular care. Fig. 1B stacks the immune-inflamed and immune-desert proportions, which is legitimate: they sum to 100%.

But in a stacked bar the only segment a person can read accurately is the bottom one. Every other segment starts at a different height, so they cannot be compared.

Stack when

the total means something (100%) and the reader's target is the bottom segment

Do not stack when

middle segments must be compared across conditions. Then give each segment its own panel

Can you see the T cells (orange) increasing on the left? On the right you can, immediately — because the baselines now match.

Python Runs in your browser

import matplotlib.pyplot as plt
import viz, numpy as np

# four conditions, cell composition - stacked versus grouped
labs = ["WT+iso", "WT+aPD1", "KD+iso", "KD+aPD1"]
cells = ["cancer", "T", "myeloid", "B/NK"]
M = np.array([[74, 8, 12, 6],
              [70, 11, 13, 6],
              [66, 14, 14, 6],
              [48, 27, 17, 8]], float)

fig, ax = plt.subplots(1, 2, figsize=(10.5, 3.4))
bot = np.zeros(4)
for j, c in enumerate(cells):
    ax[0].bar(labs, M[:, j], bottom=bot, label=c, color=viz.PAL[j])
    bot += M[:, j]
ax[0].set_title("stacked", fontsize=9); ax[0].legend(fontsize=6.5, frameon=False)

w = 0.2
for j, c in enumerate(cells):
    ax[1].bar(np.arange(4) + j*w - 1.5*w, M[:, j], w, label=c, color=viz.PAL[j])
ax[1].set_xticks(range(4)); ax[1].set_xticklabels(labs)
ax[1].set_title("grouped", fontsize=9)

for a in ax:
    a.tick_params(axis="x", rotation=30, labelsize=7)
    a.set_ylabel("% of cells", fontsize=8); viz.bare(a)
fig.tight_layout()
print("These numbers only imitate the shape of Fig. 5D - they are an illustrative example.")

🤔 Check yourself

Thirteen tumour means are drawn in alphabetical order. Why is that poor?

🤔 Check yourself

Which segment of a stacked bar can people compare accurately?