DEV Community

Cover image for Root Cause Analysis Charts: How to Find Why a Metric Dropped
andreykh for AnyChart

Posted on Originally published at anychart.com

Root Cause Analysis Charts: How to Find Why a Metric Dropped

Root cause analysis leans on a different set of charts than reporting does. Seven do most of the work, and each one answers a single diagnostic question: which segment owns the drop, which few causes matter, what moved the number, where the volume leaks. None of them is a reporting chart.

That distinction is the whole problem. Most dashboards are built to report, and they are good at it. Revenue is down eight percent. Resolution time went from six hours to eleven. There it is, in red, updated hourly. Then someone asks why, and the dashboard has nothing left to say. The next twenty minutes go into filtered exports and a pivot table, and the answer that eventually turns up is one person's hunch with a number attached.

Read each one below by the question it answers. That is what decides which you reach for, not the shape of your data.

Shortlist: Which Chart Answers Which Question

Chart Type Question
Decomposition tree Which part of the number is dragging it down?
Pareto chart Of all the causes, which few actually matter?
Waterfall chart What moved the number from then to now?
Heatmap chart Which combination of two things is the problem?
Box plot What is the average hiding?
Scatter plot Do these two things really move together?
Sankey diagram Where does the volume go, and where does it leak?

Seven Charts, Seven Questions

Decomposition Tree: Which part of the number is dragging it down?

Decomposition tree breaking profit down through four dimensions in sequence: product category, then city, then ship mode, then customer type, following the loss-making branch at each level.

A decomposition tree breaks a measure down level by level, across as many dimensions as you need, with each split chosen as you go.

Reach for it when a KPI is off and you don't yet know which segment owns it. It is usually the first move, because it narrows the search space before you have committed to a theory.

Watch out: The biggest branch feels like the answer. It's only where to start looking. A branch can be large simply because that segment is large.

Cousins: Treemap and Sunburst show the same breakdown, but on a fixed hierarchy with no live drill. Power BI ships a decomposition tree natively; in Qlik Sense it arrives as an extension.

Pareto Chart: Of all the causes, which few actually matter?

Pareto chart of customer complaints: bars ranked from most to least frequent with a cumulative line crossing them, showing how few categories account for most complaints

A Pareto chart ranks categories largest to smallest, with a cumulative line across them.

Best when a long tail of small causes is hiding a short head of big ones. The cumulative line is doing the real work: it tells you where to stop reading.

Watch out: 80/20 is a pattern, not a law. Don't stop at the tallest bar; the 2nd and 3rd are often the ones you can move. The largest cause is frequently the one nobody has authority over.

Cousins: A sorted column chart is its base; the cumulative line makes it a Pareto.

Waterfall Chart: What moved the number from then to now?

Waterfall chart of an income statement by region, bridging a starting figure to an ending one through the positive and negative steps in between.

A waterfall chart shows the step-by-step ups and downs bridging a start value to an end value.

This is the chart for explaining a change: budget vs. actual, last period vs. this one. It is the only one here that reconstructs a number rather than dissecting it.

Watch out: Only honest for additive parts. If factors overlap or interact, the bridge misleads. Price and volume effects are the classic trap, since a change in one moves the other.

Cousins: None. It's the purpose-built variance chart.

Heatmap Chart: Which combination of two things is the problem?

Heatmap of daily climate data laid out as a grid, with color intensity making the seasonal pattern visible across two dimensions at once.

A heatmap chart puts one dimension down the rows, another across the columns, with each cell colored by the measure.

Best for finding the exact cell: product × region, day × hour, channel × segment. A decomposition tree tells you the region is down; a heatmap tells you it is that region on weekends.

Watch out: Too many rows or columns and it's just noise. Sort both axes by value, never alphabetically. Alphabetical order scatters the pattern that sorting would have made obvious in one glance.

Cousins: A pivot table is the un-encoded version; color is what surfaces the pattern.

Box Plot: What is the average hiding?

Box plot comparing happiness scores by region, showing the median, the quartile spread and the outliers for each group on one scale.

A box plot shows the median, quartiles, and outliers for each group on one scale.

Use it for groups where the spread or the outliers, not the mean, are the story. It is the fastest way to find out that a metric did not get worse across the board.

Watch out: It can hide a two-peak distribution: it reads as one wide box. When shape matters, use a histogram.

Cousins: A histogram and a distribution plot show the full shape of a single distribution.

Scatter Plot: Do these two things really move together?

Scatter plot with a fitted trend channel, showing two measures against each other. A tight cloud like this is a lead to test, not proof of causation.

A pattern this clean still does not tell you why.

A scatter plot, or dot chart, puts two measures on X and Y, one dot per record.

The job here is pressure-testing a suspected driver before you trust it. By this point in an investigation you usually have a theory, and this is the chart that tries to break it.

Watch out: Correlation is never causation. A tight cloud is a lead, not proof; watch for a hidden third variable. This is the single most common way a root cause analysis produces a confident wrong answer.

Cousins: A bubble chart adds a third variable as size; a scatter matrix shows many pairs at once.

Sankey Diagram: Where does the volume go, and where does it leak?

Sankey diagram of a user flow, with link widths sized to volume so that the paths carrying the most traffic and the points where it drops off are visible.

A Sankey diagram draws flows between stages, with link width sized to volume.

Best for following a path end to end: traffic → conversion, intake → resolution. Where the other six examine a number, this one examines a route.

Watch out: Too many nodes and it turns to spaghetti. Best with a clear direction and a few stages.

Cousins: A funnel chart is the linear, single-path case for step-by-step drop-off.

How They Chain: A Worked Example

A shortlist makes these look like seven options to pick between. In practice, you use three or four in sequence, and each one narrows what the next has to explain.

Take a support team whose average ticket resolution time has gone from six hours to eleven over a quarter. The dashboard reports the eleven. Nothing on it says why.

Three-step root cause chain: a decomposition tree rules out a segment-specific cause, a box plot shows a small slow tail dragging the average, and a Pareto ranks that tail down to two categories.

The same investigation as a map: three charts, and what each one rules out.

Start with the decomposition tree. Split resolution time by channel, then by region, then by ticket type. Suppose the split comes back nearly even: every branch is up, no single segment owns it. That is a useful answer, and it is one most investigations never get, because they stop at the first branch that looks large and start building a story around it. An even split rules out a segment-specific cause in the dimensions you checked and sends you to a different question.

Then the box plot. Group the tickets by month and look at the spread instead of the mean. Suppose the median has barely moved, from five hours to five and a half, while the upper whisker has stretched from twelve hours to forty. Now the picture is different. Most tickets are being handled at roughly the old speed. A small number are taking dramatically longer, and they are dragging the average with them. The eleven-hour figure on the dashboard describes almost none of the actual tickets.

This is why the box plot earns its place. The mean was never wrong, it was just answering a question nobody asked.

Then the Pareto. Take only the slow tail and rank it by category. If two categories account for most of the added hours, the investigation has a target. If the added hours are spread evenly across fifteen categories, it has a process problem rather than a category problem, and the next question is about routing rather than about content.

Three charts, three narrowing steps, and at no point did anyone need a hunch. The chain also shows why the order matters: running the Pareto first would have ranked all tickets and buried the slow tail inside the volume of normal ones.

What Is Not Here

You won't find gauges or bullet charts.

They monitor a number. Neither of them helps you find a cause. A gauge can tell you resolution time is in the red band. It cannot tell you that the red is coming from forty tickets out of nine hundred.

That is not a criticism of those charts. They are built for a different job: telling you whether a number is where it should be. The mistake is reaching for a monitoring chart when the question has moved from what to why.

Two charts on this list are worth naming for what they cannot do. A scatter plot shows correlation and never causation, no matter how tight the cloud. And a waterfall is only honest when the factors it bridges are genuinely additive. Both produce confident, well-formatted, wrong answers when used past their limits, which makes them more dangerous than a chart that simply looks bad.

What to Do With This

If you build dashboards, the practical takeaway is smaller than a redesign. Pick the one metric people most often ask "but why" about, and add the chart that answers the second question.

Reporting the number and diagnosing it are two different jobs, and most dashboards have only ever been asked to do the first.

These seven answer why a number moved. Whether it is where it should be is a different job, one that performance vs. target charts are built for.

Top comments (0)