Show Uncertainty in a Chart Without Drawing a Confidence Interval You Did Not Calculate
Show Uncertainty in a Chart Without Drawing a Confidence Interval You Did Not Calculate
You have a number you don't fully trust. You also have a shape — the soft shaded band — that makes a chart look careful. The temptation is to put the second thing on the first and call it honesty.
It usually isn't. A band is a shape. It has no memory of where it came from.
A range built from three planning assumptions, a range calculated from sample data, and a range that was never estimated at all can be drawn to look identical down to the pixel. Strip off the labels and a reader can't tell them apart. That's the problem, and it isn't solved by picking a nicer shade of gray. It's solved by knowing what produced your endpoints and labeling the chart so that a reader knows too.
This is not an article about calculating an interval. If you need a real confidence interval and don't have one, nothing here will help you produce it — that's a job for a method and someone qualified to apply it. This is about the step before you draw: working out what kind of uncertainty you actually have, and refusing to draw the kind you don't.
Trace the uncertain value back to its source
Before you touch the chart, find out what the number is. Not what it says — what it is. There are at least four different things that arrive at a desk wearing the same "uncertain" label, and they need different treatment.
Measured variation. You have repeated observations of something: weekly signups, test scores, sensor readings, delivery times. The number wobbles because the thing itself wobbles. If you have a method for turning that wobble into an interval, and someone qualified has applied it, you have a real statistical interval. If you only have the numbers, you have a spread you can describe but not an interval you can calculate from scratch.
A supplied estimate with a calculated interval. Someone else did the statistics. This is the good case, and it comes with obligations rather than freedom — the parameter, the method, and the coverage level belong to whoever made it, not to you. More on this below.
Alternative assumptions. No statistics were involved. Someone decided to plan against two or three named futures: a slow-intake case, a planned case, a stretch case. The spread you see is the spread of the assumptions that were chosen. It is a useful and honest thing to chart. It is not a probability distribution, and it is not a statistical interval.
An unresolved input. Nobody has estimated the thing yet. There is no method, no data, no assumption — just a question that hasn't been answered. There is no range here at all.
These get mixed together constantly, which is why the mixing has to be sorted out before design. A central value with a "+/-" from a finance model, a range of outcomes from a pilot, and a hope from a sales conversation can all end up in one spreadsheet column called "low, mid, high." Separating them is not a formatting task. It's the actual work.
Ask who supplied the central number and what they meant by it. Often the answer tells you which of the four you have. If someone says "that's our statistical estimate" and can name the method, you're in case two. If someone says "those are just the numbers we're planning against," you're in case three. If someone says "we haven't sized it yet," you're in case four, and the correct chart element may turn out to be a sentence.
A designer should not resolve an absent method by choosing a familiar-looking visual. The shape is not an answer to a question about provenance.
What the endpoints do and do not mean
Take any boundary you're about to draw and finish this sentence:
The upper end is ___ because ___.
That's the whole test, and it will keep you out of most of the trouble in this article. It forces you to give the endpoint a reason rather than a position.
For a set of scenarios, the reason is an assumption: "The upper end is 120 because it assumes the paid channel in testing works at the rate we've seen so far, on top of the two launches already scheduled." Notice what that sentence does not say. It does not say 120 is the highest possible outcome. It doesn't say there's a particular chance of landing near it. Someone could absolutely sign up faster than that — the arithmetic doesn't stop you. A scenario range is the span of futures you chose to consider, and the edge of what you considered is not the edge of what can happen.
That distinction is worth an explicit label, because it's the one readers get wrong most often. "Range of planning cases" or "Assumption range" tells the truth. "Confidence range" tells a lie with a familiar word in it.
For a supplied statistical interval, the sentence has a different shape: "This interval covers the true value in 95% of samples drawn this way, under the assumptions in the source's method note." The reason is a procedure, not a preference. Four things need to travel with it intact — what quantity is being estimated, how the interval was produced, what the coverage level means in the source's own words, and the source and date. If you can't carry all four, you can carry the source's own wording next to the chart. Paraphrasing an interval into something punchier is how a defensible estimate becomes an indefensible claim.
One more trap, and it's the quiet one. A minimum and a maximum taken from a set of observations — the fastest and slowest delivery, the cheapest and most expensive quote — is a description of what you saw. It is not a confidence interval, and it doesn't cover anything beyond the observations themselves. It looks statistical, and it isn't. Label it "observed range" and don't let the styling imply more.
Three panels that look the same and aren't
Here's the comparison. Three charts, all of which someone in the meeting would happily describe as "the range."
Panel one: a scenario range. Fictional, and deliberately so: a small company is planning weekly request volume for the coming year against three named intake assumptions.
- Low intake — 80 requests per week. Signups continue at the rate of the last two months, with no new channels.
- Planned intake — 100 requests per week. The two launches already scheduled add volume at the rate they were planned to.
- High intake — 120 requests per week. Planned, plus the paid channel currently in testing, if it holds up.
Three named cases, three lines or three bars, each labeled with the assumption behind it. No probabilities attached. No statement that the middle case is more likely. No subtitle reading "80–120." And critically, no filled envelope between 80 and 120 unless its edges are labeled with the assumption names, because an unlabeled envelope looks like a distribution, and this isn't one.
The endpoint sentence works here: the upper end is 120 because of a named assumption about an untested channel. Honest, useful, checkable. If that channel doesn't work, the sentence stops being true and the chart changes — which is exactly the behavior you want from a planning chart.
Panel two: a supplied interval. This article doesn't carry a numerical example for this panel, because none was verified while it was being written. That absence is the point, so let's use it.
Here is the panel with its slots named rather than filled:
- Parameter: what quantity the interval estimates.
- Method: how it was produced.
- Coverage level, in the source's words: what the number means.
- Source and date: where it came from.
What would be wrong here is pinning a plausible figure — a tidy "+/-4 points," say — so the panel looks finished. That would be precisely the failure this article is about, performed inside the article about it. If you're building this panel for a real deck, the numbers come from the source, entered without translation. When someone asks why the chart says what it says, the answer is a document, not a shrug.
Panel three: not estimated. A proposal to enter a new market. The deck has a revenue line for it. Someone wants a band around the line, because every other chart in the deck has one.
Nobody has estimated demand. There's no comparable market analyzed, no pricing assumption tested, no penetration rate, no period over which any of it would apply. There is no range. What belongs in the chart is a note: Demand not estimated. Beside it, if there's room, what would be needed — a comparable market with data, a price, a time period, and a method for combining them into a figure someone is willing to defend.
A flat line with no band reads as certainty. A flat line with a band reads as precision. A line that says "not estimated" reads as unknown, which is the only one of the three that's true.
Now render all three the naive way — each as an unlabeled shaded band between two endpoints — and set them side by side. Cover the labels. They are indistinguishable. Three completely different evidential situations, drawn with the same envelope: one built from choices, one inherited from a method, one with nothing behind it. In panel three, even that envelope would have to be invented, because there are no endpoints to enclose. This is why the shape alone can never be the message. The label is doing all the work, which means the label has to be built to carry it.
And it's also why you shouldn't put all three on one chart. If you merged these into a single "uncertainty plot," you'd be implying that the three quantities are comparable on the same footing. They aren't. Two are backed by different kinds of reasoning and one is backed by nothing yet. A chart that treats them alike invents a comparison that doesn't exist.
Choose a form that preserves the distinction
Once you know what you have, the form follows — with a few specific choices worth making deliberately.
Scenarios generally want separate, labeled marks. Three lines or three bars, each named. A filled band between the outer cases invites readers to picture density, with more likelihood in the middle and less at the edges. If your cases are discrete alternatives, that picture is wrong. Separate marks avoid implying it.
A properly sourced interval can use a range display, drawn between the endpoints its source gave you, with the meaning nearby — in the axis label, the subtitle, or a source line the reader can actually reach without hunting.
Missing knowledge often belongs in a statement, not on an axis. Inventing a y-axis that accommodates a range you don't have is a way of granting it a place in the argument it hasn't earned.
One thing worth saying plainly, because it's easy to assume otherwise: not every chart needs an uncertainty band. UK statistical-presentation guidance from the Office for National Statistics, which covers showing uncertainty in charts, treats the decision to display a range as a question about how the chart will be read, and calls for the range to be explained to the reader. That's presentation guidance — it is not a method for constructing a confidence interval out of rough inputs, and its conventions don't transfer into a statistical claim you haven't earned. (The page is at https://service-manual.ons.gov.uk/data-visualisation/guidance/showing-uncertainty-in-charts; the example charts it links weren't examined for this piece.)
Adding a band to a chart that doesn't need one isn't neutral. It costs attention and it implies a question was asked and answered.
Check what the headline claims about the range
The chart is only half the claim. The other half is the sentence above it.
"We'll land between 80 and 120." Says the outer cases are bounds on what can happen. They're bounds on what was considered. Rewrite it as the cases: "Across three intake assumptions, we plan for 80 to 120 requests per week."
"Up to 120 by Q4." Takes the top of a scenario set and presents it as a ceiling. If the high case assumed something untested, this sentence quietly converts a hopeful assumption into a forecast.
"Well above last year." Where does the gap come from? If the comparison depends on a boundary whose meaning is unclear, the claim inherits that unclarity. Fix the boundary or weaken the sentence.
Two intervals that don't overlap. This reads to an audience as "significantly different," and it is not a significance test. Two estimates with non-overlapping ranges may or may not differ in any formal sense, and the chart can't tell you which. Don't let the visual separation do statistical work.
The scale. A range that matters at the scale of the decision has to be visible at the scale of the chart. A band of a few percent on an axis starting at zero is invisible; the same band on a truncated axis looks enormous. Both choices change the argument. Make the choice on purpose, and if you truncate the axis to make a consequential range legible, say so, because the bar heights are now carrying a different message than they were before.
The reviewer. Any statistical interpretation on the chart — a coverage level, a claim of significance, the word "confidence" itself — should be checked by someone qualified to check it before the deck goes out. Not after the meeting, when a number has already been repeated twice.
Keep the source note short and keep it where the claim is. A methodology paragraph on slide fourteen doesn't help someone looking at slide four.
Write the sentence first
Go back to the sentence. The upper end is ___ because ___.
If you can finish it — "because the paid channel holds at the tested rate," "because the survey's method note defines the interval this way," "because the source's method note says so and I have it in front of me" — then draw the band, label it with that reason, and put the reason somewhere the reader will find it.
If the best you have is that it looks about right, or that every other chart has one, or that someone will ask why there's no range, then you don't have a range. You have reassurance wearing the shape of a range, and the shape is the part that lies.
Write the endpoint sentence. If it can't be written accurately, the band isn't ready to be drawn.
Frequently asked questions
What are the different kinds of uncertainty that can arrive at a desk, and why does the kind matter?
At least four: measured variation (repeated observations, where a real statistical interval exists only if you have a method and a qualified person applied it); a supplied estimate with a calculated interval (where the parameter, method, and coverage level belong to whoever made it); alternative assumptions (planning against two or three named futures, not a probability distribution); and an unresolved input (nobody has estimated the thing yet, so there is no range at all). They need different treatment because they get mixed together constantly, and the mixing has to be sorted out before design.
What single sentence can test whether an endpoint is ready to draw?
Finish this sentence: "The upper end is ___ because ___." For a scenario range, the reason is an assumption—e.g., "The upper end is 120 because it assumes the paid channel in testing works at the rate we've seen so far." That does not say 120 is the highest possible outcome or that there's a particular chance of landing near it. For a supplied statistical interval, the reason is a procedure, not a preference. If you can't finish the sentence accurately, the band isn't ready to be drawn.
Why shouldn't a scenario range be drawn as a filled shaded band?
A filled band between the outer cases invites readers to picture density, with more likelihood in the middle and less at the edges. If your cases are discrete alternatives, that picture is wrong. Separate, labeled marks (three lines or three bars, each named with the assumption behind it) avoid implying a distribution. Also, an unlabeled envelope looks like a distribution, and this isn't one.
What should a chart show when a quantity has not been estimated at all?
If nobody has estimated demand—no comparable market analyzed, no pricing assumption tested, no penetration rate, no time period—then there is no range. What belongs in the chart is a note: "Demand not estimated." Beside it, if there's room, what would be needed: a comparable market with data, a price, a time period, and a method for combining them into a figure someone is willing to defend. A flat line with no band reads as certainty; a flat line with a band reads as precision; a line that says "not estimated" reads as unknown, which is the only one of the three that's true.
What headline claims about a range should trigger a rewrite?
"We'll land between 80 and 120" says the outer cases are bounds on what can happen, when they're bounds on what was considered—rewrite as the cases. "Up to 120 by Q4" takes the top of a scenario set and presents it as a ceiling. "Well above last year" inherits unclarity if the comparison depends on a boundary whose meaning is unclear. Two intervals that don't overlap read as "significantly different," which is not a significance test. And the scale matters: a range that matters at the scale of the decision must be visible at the scale of the chart; if you truncate the axis to make a consequential range legible, say so.