BitBootcampBitBootcamp ← all explainers 🔄 Review ⚙ Settings

DA-01-018-P · Primer — Choosing the Right Chart and Communicating Findings Honestly

The idea in one sentence

The right chart is the one whose shape matches the question you're asking, and an honest chart is one whose visual size matches the real size of the effect — no truncated axes, no cherry-picked slice that flips the story.

Why it exists

By now you can compute a mean, read a histogram, and score a correlation — but analysis isn't finished until someone else can believe it, and a chart is where your careful numbers meet a reader's fast, trusting eye. That eye is easy to fool, and it doesn't fool itself — the analyst does the fooling, usually by accident. Two failures dominate. First, the wrong chart: a bar chart pressed into service to show a trend, a line chart drawn over unordered categories, a pie chart splitting a "share" that doesn't sum to a whole — each one asks the reader's eye to answer a question the data can't. Second, the dishonest chart: the numbers are correct, but the drawing exaggerates them — an axis that doesn't start at zero, a y-range zoomed to make a 6% bump look like a doubling, or a start date hand-picked so a flat trend reads as a rocket. This primer exists because the last mile of analysis — presenting it — is where good work most often turns into a lie you didn't mean to tell. Choosing the chart that fits the question and drawing it without tricks is not decoration; it is the honesty of the whole analysis, made visible.

How it works

Chart choice starts from the question, not the data: "how did one number change over ordered time?" → a line chart; "how do a few categories compare in size?" → a bar chart; "how do two numbers move together?" → a scatter; "what's the shape of one column?" → a histogram. Once the chart type fits, the second job is making the ink honest: the length or height a reader sees must be proportional to the value, which for bars and most trend charts means the axis must include zero so a small change looks small. The single most common lie is the truncated y-axis — zoom the range to hug the data and a trivial wiggle balloons into a dramatic climb. So the discipline is: pick the chart from the question, then set ylim to start at zero (or to a range you can defend out loud), then check that the visual size of the change equals its real size. Here we plot six months of signups the honest way and measure how big the change actually is:

import matplotlib
matplotlib.use("Agg")            # headless backend; omit in a notebook
import pandas as pd
import matplotlib.pyplot as plt

# Monthly signups for a product, one row per ordered month
df = pd.DataFrame({
    "month":   ["Jan", "Feb", "Mar", "Apr", "May", "Jun"],
    "signups": [980, 995, 1010, 1005, 1030, 1040],
})

# QUESTION: "how have signups changed over 6 months?"
# A trend over ORDERED TIME => a LINE chart is the honest fit.
fig, ax = plt.subplots()
ax.plot(df["month"], df["signups"], marker="o")
ax.set_xlabel("month"); ax.set_ylabel("signups")
ax.set_ylim(0, 1100)             # start at 0 so the eye sees the TRUE size of the change

growth = (df["signups"].iloc[-1] / df["signups"].iloc[0] - 1) * 100
print(round(growth, 1))          # -> 6.1  (a modest +6.1% over six months)
# The honest line barely tilts up -> the picture matches the real +6.1%.
plt.show()

The number (+6.1%) and the picture (a gently rising, nearly flat line) tell the same story — which is exactly what makes the chart trustworthy.

The one gotcha

A truncated y-axis lets you report only true numbers and still tell a lie — the values are right, but the picture makes a tiny change look enormous. Take the same signups: the real growth is +6.1%, and against a zero baseline the change is only a sliver of the bar's height. Now crop the axis to start at the smallest value (980) instead of 0, and that same sliver stretches to fill the entire visible frame — the reader's eye, which measures length on screen, sees a change roughly 17× bigger than the truth:

import pandas as pd
df = pd.DataFrame({"month": ["Jan","Feb","Mar","Apr","May","Jun"],
                   "signups": [980, 995, 1010, 1005, 1030, 1040]})

lo, hi = df["signups"].min(), df["signups"].max()   # 980, 1040
honest      = (hi - lo) / hi        * 100  # change as % of a zero-based bar
misleading  = (hi - lo) / (hi - lo) * 100  # change as % of a 980-1040 cropped axis
print(round(honest, 1))       # -> 5.8   (the real, honest visual size)
print(round(misleading, 1))   # -> 100.0 (crop the axis: it now FILLS the frame)
# same true numbers, ~17x bigger on screen -> that gap is the lie.

Its twin is cherry-picking: keep the axis honest but hand-pick the slice — start the line in the one dip month so everything after looks like a boom, quote the single best week as "typical," or drop the categories that don't fit your point. Both tricks share one tell: the visual impression disagrees with the plain number. So the cure is the same discipline you already use for stats — make the picture and the number check each other. Ask "if I start the axis at zero and show all the data, does the story survive?" If a dramatic chart goes flat the moment you un-crop the axis or add back the months you skipped, the drama was in the drawing, not the data — and honest reporting means showing the version that survives.

See it — from a question to an honest chart

flowchart TD
    A["What question<br/>am I answering?"] --> B{"What kind<br/>of question?"}
    B -->|"Change over<br/>ordered time"| C["Line chart"]
    B -->|"Compare a few<br/>categories"| D["Bar chart"]
    B -->|"Two numbers<br/>move together?"| E["Scatter plot"]
    B -->|"Shape of one<br/>column"| F["Histogram"]
    C --> G["Start y-axis at 0<br/>(no truncation)"]
    D --> G
    G --> H["Show ALL data,<br/>no cherry-picked slice"]
    H --> I{"Picture size ==<br/>real number?"}
    I -->|"Yes"| J["Honest chart —<br/>ship it"]
    I -->|"No"| K["Fix axis / slice<br/>and re-check"]

The one-line takeaway

Pick the chart from the question (line for trends, bar for comparisons, scatter for relationships), then start the axis at zero and show every point — because an honest chart is one whose visual size equals the real size of the effect.

Where you'll meet it

You'll meet chart choice in the first thirty seconds of every dashboard and every slide — someone reaches for a pie chart to show a trend or a bar chart to plot two numbers against each other, and the fix is always the same question: what am I actually asking the reader to compare? You'll meet the truncated axis the moment a stakeholder shows you a chart where results are "exploding" — check the y-axis, and half the time the explosion is a 4% change on a range that starts at 96. You'll meet cherry-picking whenever a report's story hinges on a suspiciously convenient start date or a single highlighted week; the honest move is to widen the window and show the full series. And you'll meet all of it in interviews, where "how would you visualize this?" and "what's wrong with this chart?" are standard tests of whether you present findings honestly or just persuasively. The habit to carry out of this course: what question am I asking, does this chart's shape answer it, and would starting the axis at zero — and showing all the data — change the story I'm about to tell?

Named concepts: chart selection by question, misleading axes and cherry-picking, clear honest takeaways