Data states and testing
How to design for data you haven’t seen yet: empty, missing, partial and stale data, labels and scales that hold up, and the states to test before you publish.
You check a chart before you publish it. Nobody checks every filter combination on a dashboard, or next month’s data. Design each tile for the data it might get, not only the data it has today.
Data states
Decide in advance what each tile shows in each state, and use the same wording on every tile. The tile keeps its title, and a sentence replaces the chart. Never leave the tool’s own message: it’s generic, often in the wrong language, and says nothing about why.
There are three kinds of empty, and they mean different things. A reader who can’t tell them apart draws the wrong conclusion.
| State | Means | Show |
|---|---|---|
| No data | Nothing was measured | “No data”, and why: “This programme reports twice a year. Next report due 15 January 2027.” |
| No results | There is data, but the filters exclude it | “No results for these filters”, and the way out: “Clear the region filter to see all regions.” |
| Error | A load or calculation failed | What failed and the last good refresh: “The incidents data didn’t load. Last refreshed 27 October 2026.” |
The other states:
| State | Show | For example |
|---|---|---|
| Loading | The tile’s title and an empty chart area | |
| No access | Empty, with an explanation. Never an error, and never all the data | “You don’t have access to programme data for this region.” |
| Zero | The chart, with the zero shown and labelled | A bar at 0, labelled “0” |
| Missing values | A gap in the line, a dash in a table, or a Steel “No data” category | “No data” in the legend or a note |
| Partial period | The value, with a dashed line or lighter fill | “Partial week” |
| Stale | The last good data, with a warning next to the update time | “Not updated since 21 October 2026: source delayed.” |
| Too few to show | Nothing, and the reason | “Fewer than 5 cases. Hidden to protect privacy.” |
Never show missing data as zero. A zero says nothing happened. A gap says nobody knows. A dashboard about 17 of 31 programmes that doesn’t say so is wrong.
Show a refresh warning in one place: one line under the header, or in the affected tile. The same warning twice on one screen reads as two problems.
Labels and scales that hold up
- A title that includes a filter value (“Incidents in {region}”) must still read well with the longest value, and with “All”.
- Wrap or abbreviate long category names consistently. Don’t cut them off mid-word.
- A filter that removes the largest value rescales a floating axis. Fix the axis where readers compare selections, and let it float only where the shape matters more than the size.
- Show the top categories plus “Other”, so a filter can’t produce a chart with thirty bars.
- Bars still start at zero, whatever the data. If a value axis doesn’t start at zero, the subtitle says why.
- Fix the zoom and frame of a map, so it doesn’t move with every filter. Colour areas with no data Steel on purpose: left white, they look like zero.
- Give table columns fixed widths. Automatic widths jump with every filter change, and two screenshots of the same table can’t be compared.
Testing
You check a static chart by looking at it. You check a dashboard by going through the states a reader can reach and you haven’t seen. Before you publish, check:
- The opening view, with no selection. Is every number right, and does it say which period?
- One selection at the deepest level. Are the totals still right?
- A selection with no results. Does the no-results message appear, with a way out?
- The first period in the data. A comparison with “the previous period” mustn’t fall back to zero.
- The latest period, still running. Does it say it’s incomplete?
- An item that reports late. Does it show as No data, or carry over its last value? Is that what you intended?
- A reader without access. Empty with an explanation, not an error, and certainly not everything.
- The longest names and an outlier. Do labels and scales hold up?
- A screenshot, without the filters around it. Does it still say what it shows, for when and for whom?
If one of these fails, it’s rarely a formatting problem. It’s a calculation that’s wrong at one level, or text that was typed where it should come from the data.