Data visualisation principles

Intuition

A chart is a claim about the data, made visually instead of in words — and like any claim, it can be accurate, misleading, or simply the wrong claim for the question being asked. The mechanics of a chart (which chart type, which axis scale, what gets aggregated before plotting) aren’t neutral presentation choices layered on top of “the data” — they are the analysis, expressed visually, and a chart built without deliberate choices at each of these points tends to default toward the choice that looks most dramatic rather than the one that’s most accurate.

The through-line for this whole page: every visual decision that makes a chart more visually striking without a corresponding justification in the data is a decision that makes the chart more likely to mislead, whether or not that was the intent.

Mechanics

The truncated axis, and what it actually claims

Bar chart, y-axis from 95 to 100:
  Product A: 97%
  Product B: 98%

Rendered with a y-axis starting at 95 rather than 0, Product B’s bar looks roughly twice as tall as Product A’s — visually claiming “B is substantially better than A.” The actual difference is one percentage point. Truncating the axis isn’t inherently dishonest — for data that genuinely only varies in a narrow range (say, a metric that’s always between 95% and 100%, where 0–95% would compress all the meaningful variation into a sliver), a truncated axis can be the more honest choice. The failure is truncating without a corresponding annotation or justification — the same chart with the axis clearly labeled and the choice explained is a legitimate visualization; without that, it’s a one-point difference dressed up to look like a two-fold one.

Aggregation before plotting: the decision that determines what the chart can show

# plotting the mean collapses the distribution entirely
df.groupby('region')['order_value'].mean().plot(kind='bar')

# a box plot or violin plot preserves the spread the mean hides
df.boxplot(column='order_value', by='region')

A bar chart of means answers “what’s typical, on average, per group” and actively hides everything about the shape of each group’s distribution — Data profiling covers why mean alone can misrepresent a skewed distribution; a bar chart of means inherits that exact problem, visually. Two regions with identical means and wildly different variances (one region consistently near the mean, another with huge swings above and below it) render as identical bars — the chart type itself has discarded the information that would distinguish them, before any question of styling or color even comes up.

Chart type as a claim about the relationship in the data

• A line chart implies continuity and order between points — appropriate for a genuine time series, actively misleading for categorical data with no natural ordering (connecting “Region A” to “Region B” to “Region C” with a line implies a progression between them that doesn’t exist).
• A pie chart implies the segments are parts of one meaningful whole that sums to 100% — misleading when segments overlap, or when the “whole” itself isn’t the thing anyone actually cares about (individual segment comparison is usually the real question, and a pie chart is one of the weakest chart types for comparing individual values against each other).
• A dual-axis chart (two different y-axes on the same plot) implies a relationship between the two series that the shared visual space suggests but the independent scales don’t actually establish — two unrelated metrics plotted together, each on a scale chosen to make them visually correlate, is a well-known way to imply causation or correlation that isn’t in the underlying numbers at all.

When a chart is the wrong tool entirely

A single precise number — “conversion rate this month: 3.2%, versus 3.1% last month” — is often better communicated as text or a simple table than as a chart, because a chart’s strength is showing patterns across many data points (a trend, a distribution, a comparison across several categories), and forcing a two-number comparison into chart form usually adds visual overhead without adding clarity. The decision to chart something at all should follow from “does this have enough data points, or enough structure, for a visual pattern to matter” — not from a default assumption that data always benefits from being charted.

Cost & limits

A dashboard with many charts costs the viewer’s attention, not just engineering time to build — every additional chart competes for a viewer’s limited attention and interpretation effort, and a dashboard with twenty charts where three would answer the actual question people come to it for isn’t more thorough, it’s more likely for the important three to be missed among the other seventeen.

Real-time or frequently-refreshing charts carry a genuine computation and rendering cost proportional to refresh frequency and query complexity behind each chart — a dashboard refreshing every few seconds, showing data that only meaningfully changes hourly, is paying a real infrastructure cost for freshness nobody needed.

When NOT to use it

Do not truncate an axis without labeling the truncation clearly enough that a viewer skimming the chart (not reading every axis label carefully) still gets an accurate impression of the actual magnitude of difference shown. If that’s not achievable for a given chart, start the axis at zero instead.

Do not use a chart type that implies a relationship (continuity via a line, proportion-of-whole via a pie) the underlying data doesn’t actually have. Categorical comparisons are bar charts; genuine time series are line charts; parts of a true, meaningful whole are (cautiously) pie charts — matching the type to the actual structure of the data is not a stylistic detail.

Do not build a chart before deciding what aggregation level answers the actual question. A chart built on whatever aggregation the data happened to arrive in (raw transaction-level data plotted directly, or an average computed by default) rather than the aggregation the question actually calls for produces a chart that’s technically accurate about the wrong thing.

Real-world usage

Executive dashboards are the highest-stakes application of these principles, because they’re often viewed briefly, by people who won’t cross-check the underlying data, making a misleading (even if unintentionally so) chart more likely to directly drive a wrong decision than the same chart buried in a detailed analytical report a careful reader would scrutinize. A/B test result reporting is a related high-stakes case, where an inappropriately truncated axis or an average-only comparison (ignoring variance) can make a statistically insignificant difference look like a clear win.

Failure modes

The dashboard that drove a decision based on a visual exaggeration. A truncated-axis chart or a dual-axis chart implying a correlation stronger than the data supports can genuinely change a business decision — the failure isn’t caught by “the chart was wrong” (the numbers plotted were usually accurate) but by “the visual impression the chart created didn’t match the actual magnitude or relationship in the data,” which is a subtler and more consequential failure than a simple data error.

The bar-chart-of-means that hid the real story. Two groups with identical average performance and wildly different consistency (one reliable, one volatile) render identically in a bar chart of means — a decision made from that chart alone (which region should we expand investment in?) misses the variance difference entirely, because the chart type discarded exactly the information that would have surfaced it.

The dashboard nobody looked at because it had too many charts. A dashboard built to be comprehensive rather than focused can fail simply by overwhelming its viewers — the specific chart that mattered for a given decision existing somewhere on the dashboard isn’t the same as it being seen and acted on, and a dashboard’s actual usefulness is measured by whether the important signal reaches the viewer, not by how much it displays.

Practice problems

1. A chart shows a y-axis from 80 to 85, making a change from 81% to 84% look like a dramatic jump. Is truncating this axis inherently wrong?

Not inherently — if the metric genuinely and meaningfully only varies within a narrow band (say, it’s never been outside 78–86% historically), a truncated axis can correctly emphasize real, meaningful variation that a 0–100 axis would flatten into invisibility. The question is whether the truncation is clearly labeled and the choice explained, versus silently exaggerating an otherwise unremarkable change.

2. A team wants to compare order value across five regions. Someone proposes a pie chart with five slices. What would you suggest instead, and why?

A bar chart — pie charts are one of the weakest chart types for comparing individual values against each other, because human perception is better at comparing bar lengths (a shared linear scale) than at comparing wedge angles or areas. A pie chart is more appropriate when the actual question is about proportion-of-a-meaningful-whole rather than individual-value comparison, which usually isn’t the real question behind “compare X across categories.”

3. A dashboard chart shows two unrelated metrics (revenue and website uptime) on a dual-axis line chart, with axes scaled so both lines appear to move together. What’s the risk, and what would you check?

Dual-axis charts with independently scaled axes can make two genuinely unrelated metrics appear correlated purely through axis-scaling choice — check the actual correlation coefficient between the two series numerically rather than trusting the visual impression, and consider whether plotting them separately (or on a shared, meaningful scale) would avoid implying a relationship the data doesn’t support.

Check yourself

A bar chart's y-axis is truncated to start at 95 instead of 0, making a one-point difference (97% vs 98%) look like roughly a two-fold difference in bar height. Is this always misleading?

• Yes, axes should always start at zero
• Not inherently -- it depends on whether the truncation is clearly labeled and justified by the data genuinely only varying in a narrow band
• No, truncated axes are always more accurate
• It only matters for line charts, not bar charts

Interview answers

“What makes a chart misleading, even if the underlying data is correct?” The chart’s visual claims — axis scale, chart type, aggregation level — can imply a magnitude or relationship in the data that the actual numbers don’t support, independent of whether any individual data point plotted is accurate. A truncated, unlabeled axis exaggerating a small difference, or a chart type implying continuity or proportion the data doesn’t have, are both cases where “every number shown is correct” and “the chart accurately represents the data” are different claims. The caveat: this isn’t an argument against ever truncating an axis or choosing a dramatic chart type — it’s an argument for the choice being deliberate and legible to the viewer, not incidental.

“How would you decide what chart type to use for a given dataset?” Start from what relationship or comparison the question actually needs — a genuine time series calls for a line chart (implying continuity correctly); categorical comparison calls for bars (which human perception compares more accurately than angles or areas); true parts-of-a-whole calls for a pie chart, cautiously. The caveat that shows real experience: the aggregation level matters as much as the chart type — a bar chart of means can visually equate two groups with identical averages and wildly different variances, discarding exactly the information that would distinguish them, before any question of chart type or styling even comes into play.