Why a chart is part of the analysis
A table of five hundred readings tells almost nothing at a glance; a plot of the same readings can reveal a drift, a gap, a cluster of odd values or a clean linear trend in seconds. Charts are therefore a tool for finding things out as well as for presenting them. The same chart that helps you understand your data can mislead a reader if it is built carelessly, so the choices you make matter.
This article gives general guidance on charting engineering data. It is not specific to any software.
Match the chart to the question
Start from what you want a reader to see.
- Relationship between two quantities: a scatter plot, with each point one observation. It shows trends, clusters and outliers without assuming a smooth link.
- Change over time: a line chart, with time on the horizontal axis. Connect points only when the quantity really changes continuously or the connection helps reading.
- Distribution of values: a histogram, which shows how readings are spread, whether they are symmetric and whether there are two separate groups. Related summaries such as the mean and standard deviation describe the same data in numbers.
- Comparison between categories: a bar chart, for example failure counts by supplier. Bars should start at zero, because their length represents the value.
- Parts of a whole: use sparingly. Pie charts become hard to read with many slices; a sorted bar chart is usually clearer.
Avoid decoration. Three-dimensional effects, heavy shading and unnecessary colour make values harder to compare accurately.
Honest axes
Axes are the easiest place to mislead without intending to.
- Truncated axes. A bar chart that starts at 95 instead of 0 makes a small difference look dramatic. For line and scatter charts a non-zero start can be reasonable, since position rather than length carries the meaning, but state clearly what range is shown.
- Different scales on a shared chart. Two lines on separate vertical axes can be made to look as though they track each other simply by adjusting the scales.
- Logarithmic axes. These are useful when data spans orders of magnitude, such as frequency responses expressed on a decibel scale, but they must be labelled, because equal distances then mean equal ratios rather than equal differences.
- Units and labels. Every axis needs a name and a unit. A plot without units is a drawing, not evidence.
A good test: if someone swapped your chart for one with different axis limits, would the story change? If so, think about which version is the fair one and say why.
Show uncertainty
Measurements are never exact, and a chart that shows only single values can imply more certainty than exists. Options include:
- Error bars showing the estimated uncertainty of each point. State what they represent, such as one standard deviation or the uncertainty of the mean; these are different things. The ideas behind measurement uncertainty explain the distinction.
- Showing all the points, for small datasets, so readers see the actual scatter rather than only an average.
- Bands around a fitted line to show the range within which the trend is plausible.
- Box plots to display median, spread and outliers across several groups.
As an illustration, imagine two designs whose average results differ by a small amount while the individual readings overlap heavily. A bar chart of the two averages would suggest a clear winner; a plot of all the points makes the overlap, and so the doubt, obvious.
Reading charts critically
When you meet a chart from someone else, ask a few questions. What are the axis ranges and units? How many data points lie behind each mark? Is the vertical axis truncated? Is a trend line fitted to a narrow range and extended beyond it? Does a correlation get presented as proof of cause? Is there a reason some data might have been left out? Checking against basic sanity checks, such as whether the magnitudes are plausible, catches many problems.
Fitted lines deserve particular care. A line drawn through data shows the trend within the measured range; extending it outside that range is extrapolation and can be badly wrong, since many physical systems stop behaving linearly.
Presenting to decision makers
Decision makers usually have limited time and may not share your technical background, so design for them.
- Lead with the message. Give the chart a title that states the finding, such as “Failure rate fell after the change”, rather than a bare label.
- One main point per chart. If you need several messages, use several charts.
- Highlight what matters, perhaps with a limit line or a single emphasised series, and keep the rest muted.
- Use consistent scales and colours across related charts so they can be compared. Avoid colour combinations that are hard to distinguish for colour-blind readers, and do not rely on colour alone.
- State the source and the caveats briefly: sample size, date range, known limitations.
- Be ready with the numbers. Have the underlying table available if someone asks.
These habits fit naturally into technical reports and technical presentations, and clear wording as in plain language for technical audiences makes the accompanying text easier to act on.
A short checklist
Before sharing a chart, confirm that the axes are labelled with units, the scale is fair, uncertainty is shown or its absence explained, the title says what to conclude and the chart can be understood without you in the room. A chart that a reader can misread is not finished, however good the data behind it.