See Accurately: Visual Choice and Chart Reading

30 min

The Job of a Chart: Answer a Question, Not Decorate

Learning objective: start from analytical question and audience; define one primary message per chart.

A colleague asks you to "visualize the quarterly numbers." If you open a spreadsheet and pick whatever chart type looks nice, you have already lost the thread: a chart is not decoration, it is the answer to a specific question for a specific audience. Before choosing a chart type, you need to know what question it answers and what the one sentence underneath it should say.

import pandas as pd

# clean: a multi-entity, multi-period indicator table (ESG / campus / operations)
df = pd.read_csv("DS103_Indicators_Clean.csv")
df.shape
df.head()

# Before plotting anything: what is the ONE question this chart must answer?
question = "Which entities had the highest emissions intensity in the latest period?"

• State the question in one sentence, and make it exactly one of: comparison, trend, distribution, relationship, or composition
• Name the audience as a specific role or decision-maker, not "everyone"
• Write the message so the reader can say it out loud within five seconds, with no extra explanation needed
• Write the brief before drawing anything, not as a caption added after the chart is already built

Key takeaway: A chart earns its place by answering one specific question for one specific audience, not by displaying more data.

40 min

Comparison, Trend, Distribution, Relationship, Composition

Learning objective: match analytical intent to chart family; avoid chart types that distort comparisons.

Most analytical questions fall into a small number of families, and each family has a chart type that answers it well and several that quietly distort it. Learning to name the family first makes the chart choice almost automatic.

import matplotlib.pyplot as plt

latest = df[df["period"] == df["period"].max()]
comparison = latest[latest["metric"] == "emissions_intensity"].sort_values("value", ascending=False)

# Comparison question: bar chart, sorted, shared baseline
plt.bar(comparison["entity_id"], comparison["value"])

trend = df[(df["entity_id"] == "E01") & (df["metric"] == "emissions_intensity")].sort_values("period")
# Trend question: line chart over the same metric, one entity, many periods
plt.plot(trend["period"], trend["value"])

• Name the question family before opening any charting tool: comparison, trend, distribution, relationship, or composition
• A sorted bar chart with a shared baseline is the safest default for category comparisons
• Reserve line charts for values that change over a continuous sequence, usually time
• Do not force a composition chart (pie, stacked bar) onto a comparison question; it hides the ranking the reader actually needs

1. 1Name the question family
   State whether this is a comparison, trend, distribution, relationship, or composition question before doing anything else.
2. 2Check the category count
   For comparison or composition questions, once categories run past five to seven, a bar chart is usually clearer than a pie chart or a crowded legend.
3. 3Pick the matching chart and check for distortion risk
   Use the table to choose the matching chart family, then check it against the common distortions to avoid.

Key takeaway: Naming the question family first - comparison, trend, distribution, relationship, or composition - nearly determines the right chart type on its own.

35 min

Axes, Scales, Ordering, and Baselines

Learning objective: use scales honestly; understand zero baselines, log scales, sorting, and denominator effects.

A chart can use the right chart type and still mislead through its scale. The most common ways this happens are a bar chart that does not start at zero, a linear scale used for values that span several orders of magnitude, an arbitrary sort order, and a raw count compared without its denominator.

# A bar chart's y-axis must include 0, or the visual gap is exaggerated
plt.bar(comparison["entity_id"], comparison["value"])
plt.ylim(bottom=0)

# Comparing entities of different sizes: use a ratio, not a raw count, when one exists
per_unit = comparison.assign(rate=comparison["value"] / comparison["denominator"])

• Does the bar chart's value axis start at 0?
• Do the values span multiple orders of magnitude, requiring a labeled log scale?
• Is the sort order justified by something other than favoring one conclusion?
• Has a raw count been converted to a rate before comparing entities of different sizes?

Key takeaway: A chart can use the right chart type and still mislead through its axis baseline, its scale, its sort order, or a missing denominator.

35 min

Color, Labels, and Accessibility

Learning objective: use restrained encodings; direct labels; color-independent meaning; accessible text alternatives.

Color is one of the weakest channels for carrying precise meaning: roughly one in twelve men has some form of color-vision deficiency, printed or exported charts often lose color entirely, and a legend forces the reader to look back and forth between the chart and a key. Restrained, redundant encoding fixes all three problems at once.

for _, row in trend.tail(1).iterrows():
    plt.annotate(row["entity_id"], (row["period"], row["value"]))

plt.title("Entity E01: emissions intensity fell 18% since 2023")  # states the finding, not just the axis names

• Never let color be the only channel that carries meaning; pair it with a label, shape, or pattern
• Prefer direct labels on the chart over a legend the reader must decode separately
• Write titles that state the finding or the question, not just the variable names already on the axes
• Provide a one- or two-sentence text summary of the chart's main point for screen-reader users

1. 1Preview in grayscale
   Convert the chart to grayscale or a black-and-white print preview and confirm every category is still distinguishable.
2. 2Hide the legend, read the direct labels
   Temporarily hide the legend and confirm the reader can still follow each series from its direct label alone.
3. 3Read only the title and summary aloud
   Read only the chart title and a one-sentence summary aloud and confirm a screen-reader user would still get the point.

Key takeaway: Restrained, redundant encoding - direct labels, shapes, and informative titles - makes a chart accessible instead of merely colorful.

45 min

Guided Lab: From Table to Three Evidence Charts

Learning objective: with instructor guidance, create comparison, trend, and distribution charts from a common dataset and critique them.

• Go to the DS 103 Practice page
• PreviewDownload the clean indicators file (CSV, synthetic data)
• Download the Guided Lab starter notebook
• PreviewDownload the data dictionary (Traditional Chinese)
• PreviewDownload the data dictionary (English)

1. 1Load and inspect
   Load the clean indicators file and check its shape, entities, periods, and metrics.
2. 2Build a comparison chart
   Pick one metric and the latest period; build a sorted bar chart comparing all entities, with a zero baseline.
3. 3Build a trend chart
   Pick one entity and metric; build a line chart across all available periods.
4. 4Build a distribution chart
   Pick one metric across all entities and periods; build a histogram of its values.
5. 5Write one evidence statement per chart
   For each chart, write a one-sentence claim that stays inside what that chart alone can support.
6. 6Critique each chart
   For each chart, name one real question it cannot answer on its own.
7. 7Check for a denominator effect
   Confirm whether any comparison should use a rate instead of a raw value.

• Stopping after one chart type instead of cross-checking the same conclusion with the other two families
• Writing an evidence statement that claims more than the chart it sits next to can support
• Forgetting to check for a denominator effect before comparing raw values across entities of different sizes

Key takeaway: Comparison, trend, and distribution charts each answer a different question from the same data; a trustworthy analysis uses more than one and states clearly what each can and cannot support.