Choosing the Right Chart for the Metric: A Taxonomy for People Who Ship Dashboards

Chart selection is treated as an aesthetic preference, which is why so many dashboards contain a pie chart of twelve categories and a line chart of unrelated monthly totals. It isn't aesthetic. The mapping from data shape to chart type is nearly deterministic, and getting it wrong doesn't just look bad — it hides the pattern the data was trying to show you. Here's the taxonomy we apply.

Start With the Question, Not the Data

Every chart answers one of four questions. Identify which one you're asking and the chart mostly chooses itself:

• How does it change over time? ? trend charts
• How are values spread out? ? distribution charts
• What are the parts of the whole? ? composition charts
• Which is bigger? ? comparison charts

Trends: The Line Chart Owns This Category

Time on the x-axis, metric on the y-axis, one line per series. The line chart wins because slope is the most accurately perceived visual encoding — the eye reads "accelerating," "flattening," and "inflection point" directly off a line, while the same data as bars forces you to compare bar tops and reconstruct the trend in your head. Use bars over time only when the point is individual period magnitudes (monthly revenue as discrete events) rather than the trajectory. Two rules: never use a line chart for non-ordered categories (a line connecting "North, South, East, West" implies a continuum that doesn't exist), and don't smooth a line unless you say so — smoothing is an editorial act that can erase the spike that mattered.

Distributions: Histograms, and the Bar Chart Confusion

A histogram shows how a continuous variable spreads: order values bucketed into ranges, with bar height as count. It's the chart that answers "are we normal, skewed, or bimodal?" — the chart that would have revealed the trimodal store revenue from our clustering post instantly. The common crime: calling a categorical bar chart a histogram. Bars of "sales by region" are not a histogram; regions are categories, not buckets of a number. The visual difference — histogram bars touch, category bars have gaps — encodes a real semantic difference. When the question is "where do most values sit and how fat are the tails," nothing substitutes for the histogram.

Composition: Stacked Bars, and the Pie Chart Ceasefire

Pie charts are defensible under exactly one condition: two or three slices with a takeaway like "60/40." Beyond that, humans compare angles badly — the difference between a 14% slice and a 17% slice is invisible without labels, at which point the chart is a table with extra steps. The upgrade path: a single 100% stacked bar for one whole (share by channel), and stacked bars over time when composition itself is the trend. If you need exact values for more than ~5 parts, use a sorted table. Composition charts saturate fast; treat them as summary devices, not analysis devices.

Comparisons: Bars, Sorted, Horizontal When Needed

Comparing magnitudes across categories is the bar chart's home turf — length is the most precise encoding after position. Three rules that fix 90% of bad bar charts: sort the bars (alphabetical order is never the informative order), start the axis at zero (bar length is proportional to value; a truncated axis turns a 5% difference into a visual 300% one), and go horizontal when category names are long or numerous.

The Chart Crimes List

1. Dual y-axes: two trends on different scales, inviting the reader to see a relationship in whatever vertical alignment the axes happen to create. Rescale and facet instead.
2. Truncated axes on bars: fine for lines (sometimes), fatal for bars, where length is the data.
3. 3D anything: perspective distortion makes the front slice of a 3D pie look 30% larger. It's decoration that corrupts the data.
4. Overloaded rainbow palettes: color should encode meaning (this series vs. that series, good vs. bad), not enumerate categories. Twelve colors is a legend-lookup chore, not a visualization.
5. Line charts of cumulative totals presented as performance: cumulative curves only ever go up and to the right. They look like growth even when the underlying rate is collapsing. Chart the rate, not the running total.

The One-Minute Checklist

Before shipping any chart: What question does it answer? Is the encoding the strongest one for that question (position > length > angle > color)? Does the axis start where it should? Could a table do this better? Chart choice is data-integrity work disguised as design — treat it with the same rigor as the aggregation underneath it, and your dashboards will stop lying by accident.