Notched Box Plot: What It Is and When to Use It

A notched box plot extends the classic box plot by adding a narrowing cut around the median that represents a 95 percent confidence interval. This article covers what the notch means, when to add one, how to read it correctly, and the missteps that can lead analysts to the wrong conclusion.

Key takeaways for notched box plots

If you only remember a few things, remember these:

• The notch shows uncertainty: It represents a rough 95 percent confidence interval around the median, calculated as median ± 1.58 × IQR / √n.
• Non-overlapping notches signal a difference: If notches between two groups don't overlap, you have visual evidence that medians differ at roughly the 95 percent confidence level.
• Overlapping notches don't prove equality: They only indicate insufficient evidence to claim a difference. A formal test might still find one.
• Sample size matters: You need at least 15 observations per group for the notch to be reliable. Below that, the visual becomes misleading.
• Dashboards need consistency: Notches only help if every group uses the same metric definition, time window, and inclusion rules. Otherwise you end up comparing apples to, well, other apples that were measured differently.

What is a notched box plot

Picture a standard box plot with a narrowing cut into the box on either side of the median line. That narrowing (the notch) represents an approximate confidence interval around the median.

You already know what a box plot shows. The box spans from Q1 to Q3, which is the interquartile range, or IQR. The line inside marks the median. Whiskers extend outward, and outliers appear as individual points.

The notch adds something new. It tells you how confident you can be about where the median actually sits. The standard formula is median ± 1.58 × IQR / √n, which corresponds roughly to a 95 percent confidence level.

Here is where people get confused. The IQR shows spread, how dispersed the middle 50 percent of your data is. The notch shows uncertainty, how precisely you have pinpointed the median. These are different things. Mixing them up leads to bad interpretations. A wide IQR with a narrow notch means your data varies a lot, but you've still nailed down the median with confidence. A narrow IQR with a wide notch? The opposite.

John Tukey introduced this visual in the 1970s during his work on exploratory data analysis. It remains useful for quick visual inference when you're comparing groups, especially when a simple average would hide what's going on.

When to use a notched box plot

Should you add notches to your box plot, or will they just add clutter?

Add notches when you're comparing medians across two or more groups, each group has at least 15 observations, and the distributions are roughly symmetric. Skip them when you're exploring a single distribution, when sample sizes are small or unequal, or when your audience has never seen a notched box plot before.

This comes up a lot for data analysts and BI specialists who need to explain performance differences to stakeholders. It also shows up more when data literacy is low. For example, poor data literacy ranks among the top 5 barriers to analytics success. That ranking matters because it explains why visual confidence indicators (like notches) often communicate more effectively than statistical footnotes buried in appendices. A notched box plot can act like a confidence indicator in the chart itself, which beats a slide full of footnotes.

The notch formula assumes certain conditions. When those conditions break, the notch misleads. With fewer than 10 observations, the notch can extend beyond the box itself. It looks odd and signals that the interval is not reliable.

A product team once compared conversion rates across three app variants using notched box plots with fewer than 10 observations per group. The notches overlapped, so they concluded no difference existed. A follow-up test suited to proportions (for example, a two-proportion test or a permutation test) found a statistically significant gap they had dismissed. They shipped a weaker variant than they needed to.

Data requirements for notched box plots

If you want notched box plots to hold up in a dashboard review (or an executive asks you to defend the decision), the chart needs the right inputs. This is also where data engineers get pulled in, because the visualization is only as clean as the pipeline feeding it.

At minimum, make sure each group you're plotting has:

1. Enough observations: n per group, ideally 15 or more.
2. A consistent metric definition: the same filters, time window, and grain across every group.
3. Quartiles and IQR: Q1, median, Q3, and IQR (Q3 − Q1).
4. Your whisker rule: many tools use the 1.5 × IQR rule, but not all do.
5. A clear missing data policy: decide whether you drop nulls, impute, or exclude the group.

If your BI environment has a semantic layer or reusable metrics, use it. You want one approved definition of "conversion rate" or "session duration," not five slightly different ones floating around different teams.

On the data preparation side, it helps when your transformation layer can handle quartiles, sample size aggregations, and outlier flags without a pile of custom scripts. For example, in Domo, data engineers often use Magic Transform to standardize these calculations. It supports both SQL (Structured Query Language) and no-code workflows, so analysts can build notched box plots on governed datasets instead of exported files.

How notched box plots work and how to read them

Many people assume that if two notches overlap, the medians are statistically indistinguishable. Not quite right.

Overlapping notches mean you lack strong visual evidence of a difference. Non-overlapping notches provide evidence that medians differ at roughly the 95 percent confidence level. One is a lack of evidence for a difference. The other is evidence for a difference. And honestly, that distinction trips up a lot of analysts.

Read this chart in the right order

Start by locating the medians, the horizontal lines inside each box. These are the values you're comparing.

Check the notches next. Do they overlap between groups? If not, you have visual evidence of a median difference.

Then assess the box heights. The IQR tells you about spread. A tall box means high variability in the middle 50 percent.

Look at whisker lengths. Asymmetric whiskers suggest skew. If one whisker is much longer than the other, the distribution isn't symmetric, and the notch becomes less reliable.

Finally, count outliers. Points beyond the whiskers may indicate heavy tails or data quality issues.

Common misreads to avoid

Here are the interpretations that cause the most trouble in reviews.

• Overlapping notches mean no difference: Wrong. They mean insufficient evidence. A formal test might still find a significant gap.
• Non-overlapping notches mean definitive proof: The notch is an approximation, not a substitute for a hypothesis test.
• Ignoring sample size: Notches from five observations aren't comparable to notches from 500. Always check group sizes.
• Assuming symmetric distributions: The formula assumes rough symmetry. Wildly unequal whiskers should make you skeptical.

Notched box plot variants and related charts

Not every box plot needs notches. Not every distribution comparison needs a box plot at all.

Notched vs classic box plot

The classic box plot shows distribution shape through the median, spread, and outliers. It doesn't encode uncertainty about the median.

If you're asking whether these medians are different, add notches. If you're asking what this distribution looks like, skip them.

Variable width and letter value plots

Variable width box plots encode sample size into the box width. Wider boxes mean more observations. This helps when comparing groups with very different sample sizes because it reminds viewers that narrow boxes are less reliable.

Letter value plots show additional quantiles beyond the quartiles. They're useful for large datasets where the standard box plot hides too much detail in the tails.

When to switch charts entirely

If your main question isn't "are the medians different," one of these charts will usually communicate better.

• Violin plots: Show the full distribution shape. Better when you care about bimodality or skew, not just the median.
• Strip plots or dot plots: Show every data point. Better for small samples where summary statistics hide important variation.
• Estimation plots: Show effect sizes with confidence intervals. Better when you need rigorous inference rather than a visual heuristic.

Best practices for notched box plots

A notched box plot only works if your audience can interpret it correctly, and poorly designed visualizations can mislead interpretation and hinder decision-making. Most reporting failures come from missing labels, unstated methods, or visual clutter.

Label axes with units, not just generic terms like "value" or "group." State the sample size for each group, either in the caption or on the plot. Note the whisker rule you used. The 1.5 × IQR rule is common but not universal.

Rolling this up into an executive dashboard (finance, sales, ops, take your pick)? Don't assume everyone remembers what a notch means. A one-sentence caption often saves you from five minutes of meeting detours.

If notches extend beyond the box, acknowledge this in the caption rather than hiding it. Order groups meaningfully by median, by sample size, or by a logical sequence like time or geography.

Avoid plotting more than five or six groups on a single axis. The comparison becomes unreadable. And if you're presenting to an audience unfamiliar with notches, either explain the notation or switch to a standard box plot.

If you're supporting teams at scale, consistency becomes the whole game. IT and data leaders usually care less about the chart style and more about whether every team is pulling from the same governed dataset, with the same metric definition, so notches mean the same thing everywhere.

Here's a caption template you can copy:

"Box plots show median (center line), interquartile range (box), and 1.5 × IQR whiskers. Notches represent approximate 95 percent confidence intervals around the median (±1.58 × IQR / √n). Sample sizes: Group A (n = 24), Group B (n = 31). Non-overlapping notches suggest medians differ."

Notched box plot examples

A retail analytics team compares average order values across three regions. East has 45 observations with a median of 72 and an IQR of 31. Central has 38 observations with a median of 61 and an IQR of 26. West has 52 observations with a median of 78 and an IQR of 31.

Using the formula median ± 1.58 × IQR / √n, the team calculates notch bounds. Central's upper bound is 67.7. West's lower bound is 71.2. These don't overlap. The team has visual evidence that West's median order value exceeds Central's.

A bar chart showing means wouldn't convey this uncertainty. The notched box plot shows both the median difference and the confidence around that difference.

Now consider a product team comparing session duration across two app versions. Both groups have exactly 12 observations. The notches overlap substantially.

Does this mean the medians are equal? No. With only 12 observations, the notch formula produces wide intervals. The team shouldn't conclude there's no difference. They should gather more data or run a formal test.

How to create a notched box plot in Excel

Excel's built-in box and whisker chart doesn't support true notches. You can approximate the visual by adding error bars or overlaying shapes, but the result requires manual calculation and isn't dynamically linked to your data.

For a one-off analysis, here's how to build an approximation.

The bigger issue isn't the one-time annoyance. It is the workflow. Exporting data to separate statistical tools, manually rebuilding charts, and pasting screenshots back into decks is how tool sprawl and "shadow analytics" quietly happen. Teams often spend more time maintaining workarounds than actually analyzing the data.

Prepare the data

Start by building a small summary table for each group.

1. Organize your data with one column per group.
2. Calculate summary statistics for each group in a separate table: median, Q1, Q3, IQR, n, notch lower (median - 1.58 × IQR / √n), and notch upper (median + 1.58 × IQR / √n).
3. Validate your calculations. Q3 minus Q1 should equal IQR. The median should fall between Q1 and Q3.

Build and validate the chart

Next, build the chart and double-check Excel's calculations.

1. Insert a box and whisker chart: Insert → Chart → Box and Whisker.
2. Excel calculates quartiles and whiskers automatically. Compare Excel's values against your manual calculations to catch discrepancies.
3. To approximate notches, add error bars to the median markers or overlay rectangle shapes manually. Tedious.
4. Add axis labels, a title, and a caption stating your method and sample sizes.

Excel's box and whisker chart works for exploratory analysis but lacks native notch support and dynamic updating. For dashboards or repeated reporting, having notched box plots available directly in your BI platform (for example, Domo BI) can save time and reduce mistakes, because the chart stays connected to the governed dataset and approved metric definitions.

Limitations of notched box plots and when to use alternatives

Notched box plots are a heuristic, not a statistical test. Treat them as a screening tool that flags potential differences for follow-up.

As covered earlier, the notch gets shaky with small n and with heavy skew. When either shows up, switch to a chart that shows the full distribution or run a formal test.

Non-overlapping notches suggest a difference at roughly the 95 percent level, but this isn't equivalent to a t-test or Wilcoxon test. For high-stakes decisions, run a formal test.

And then there's the audience problem. Many stakeholders have never seen a notched box plot, and only 4 percent of organizations rate data literacy as "excellent". That statistic underscores why chart choice matters as much as chart accuracy: if your audience can't read the visual, the insight never lands. If you spend more time explaining the notation than discussing the data, you've chosen the wrong visualization.

If you do have the chart in an interactive dashboard, you can meet people halfway with guided exploration. An executive might not care about the formula. They just want to ask, "Is Region West actually different?" and drill into what's driving the distribution.

If you're building these comparisons into dashboards (and want a sanity check on when notches help vs hurt), swap notes with other practitioners. Join the Domo community and see how teams are standardizing metrics, sample-size rules, and distribution visuals.