Common mistake 1 with Heat map

A rainbow colour scale, which invents boundaries where the data is smooth. This app uses a single-hue sequential scale.

Common mistake 2 with Heat map

Reading colour as though it were precise. It supports "bigger, smaller, about the same" and nothing finer.

Common mistake 3 with Heat map

A heat map where a sorted bar chart would do. If one of the two dimensions has three levels, use bars.

Common mistake 4 with Heat map

Comparing two heat maps drawn on their own colour scales. The same colour means different values in each, and only a shared scale makes them comparable.

EconStatsCraft › Heat map

Heat map

A grid of two categories with colour for the value - the chart for a table too big to read as numbers. Region against year is the common case, and the pattern in the colour is the finding.

What it needs

Columns
Value

The intuition

A heat map puts a value at every cell of a grid and encodes it as colour. It is the right chart for a matrix - a correlation matrix, a calendar, a region-by-year table - because it shows the whole thing at once and lets pattern-finding do the work. Everything then depends on the colour scale, which is where nearly all heat maps go wrong.

When to use it

Two categorical dimensions and one measurement - region by year, industry by country, hour by day. Also for correlation matrices, where the pattern of blocks is the finding.

Where the estimator comes from

Each cell (i,j) is filled with f(v_{ij}) where f maps values to colours. The map must match the data: a SEQUENTIAL scale for values with a natural low-to-high ordering, and a DIVERGING scale, symmetric about a meaningful midpoint, for values that run negative to positive. For correlations the midpoint must be zero and the range fixed at [-1,1], or the same colour means different things in two charts. Perceptually uniform maps such as viridis are built so equal steps in value are equal perceived steps in colour, which rainbow maps are not.

What it assumes

Both dimensions have few enough levels to label. Past about thirty in either direction the labels go and the chart becomes an abstract texture.
The colour scale is linear in the value and the reader can see it. Without the scale the chart is unreadable.
Cells with several rows are aggregated, so the function chosen matters as much as the colours.
The rows and columns are ordered deliberately. A heat map with arbitrary ordering hides whatever block structure it contains, and sorting by a cluster or by a marginal total is usually what makes the pattern visible.
Missing cells are distinguishable from zero cells. Both are usually drawn pale, and they mean entirely different things.

Statistical properties

The rainbow (jet) colour map is not perceptually uniform. It creates sharp visual boundaries where the data is smooth - notably at cyan and yellow - and flattens real differences elsewhere, so it invents structure and hides other structure at once.
A diverging scale needs a meaningful midpoint. Centring a diverging map on the data mean rather than on zero makes an all-positive matrix look half negative.
Colour is among the least precise encodings. A heat map is for spotting patterns, not for reading values; annotate the cells if the numbers matter.
Row and column ORDER carries as much information as the colours. Sorting or clustering the rows can turn visual noise into obvious block structure, and an alphabetical ordering usually hides it.
About 8% of men have a red-green colour deficiency, so a red-green diverging scale is unreadable for a substantial minority. Blue-orange carries the same information safely.
A truncated colour range saturates the extremes and hides how extreme they are. Always show the scale, including its limits.

How to read the output

Check the range on the colour bar before believing a pattern. A dramatic-looking map over a range of 0.02 is showing you rounding.
Look for whole rows or columns that are uniformly dark. That is the main effect; the interesting part is a cell that breaks its row and column pattern.
Empty cells are information - that combination does not occur, which is usually a fact about data collection.
Order the rows and columns meaningfully. An alphabetical heat map hides the structure that sorting by value would reveal.
The colour scale's endpoints determine the picture as much as the data do. A scale set to the full range can flatten everything if one cell is extreme, and one set to the interquartile range exaggerates the middle.
Read the row and column margins first. Most apparent structure in a heat map is the product of the two marginal distributions rather than any interaction between them.

A worked reading

A correlation matrix drawn with the colour range fitted to the data - say -0.2 to 0.9 - looks dramatic: the weakest correlation is deep blue and the strongest deep red. Fix the range at -1 to 1, as correlations require, and the same matrix is mostly pale, which is the truth: nothing in it is near either extreme. The first chart is not mislabelled or miscalculated, and every reader will come away with a stronger impression of the relationships than the numbers support.

Symptom → diagnosis → what to do

If you see this	It means	Do this
Sharp bands in smooth data	A rainbow colour map	Use viridis or another uniform map
An all-positive matrix looking half negative	Diverging scale centred on the mean	Centre it on zero
No visible pattern	Row and column order	Cluster or sort before concluding there is none
Correlation heat maps not comparable	Free colour range per chart	Fix the range at -1 to 1
Cells all one colour	One outlier stretching the scale	Clip, or use a rank or log scale, and say so
Red-green diverging scale	Unreadable for around 8% of men	Use blue-orange

Common mistakes

A rainbow colour scale, which invents boundaries where the data is smooth. This app uses a single-hue sequential scale.
Reading colour as though it were precise. It supports "bigger, smaller, about the same" and nothing finer.
A heat map where a sorted bar chart would do. If one of the two dimensions has three levels, use bars.
Comparing two heat maps drawn on their own colour scales. The same colour means different values in each, and only a shared scale makes them comparable.

Where to read this properly

Borland and Taylor (2007), "Rainbow Color Map (Still) Considered Harmful", IEEE CG&A 27(2).
Crameri, Shephard and Heron (2020), "The Misuse of Colour in Science Communication", Nature Communications 11.
Wilke, Fundamentals of Data Visualization (2019), ch. 4 and 19.

All analyses

Data & variables

Each example is simulated from a known model, so you can check whether the estimator recovers the truth - and nothing here is anyone's private data.

Filter variables

Drag a variable into a slot on the right - or just click it.

Choose an example dataset above to begin.

Choose a transformation - it opens in the data editor, with a live preview.

Apply to

Multiply by

Columns ctrl-click for several

Group by ctrl-click for several

Within groups

Statistic

Bands

How many

Cut points

Keep rows where

Value

and

Fill with

Value

Also mark which rows were filled

Rules

Base period = 100

Which part

Percent at each tail

Threshold

Survey weight optional

Direction

Names column

Values column

Largest first

Delete them, not just count them

Read this column as

Tidy the text

New name optional

EconStatsCraft rigorous, plain-spoken econometrics

Econometrics that explains itself.

Regression, panel data, time series and the whole battery of tests. Every result is followed by a verdict you can defend.

Load a dataset to see it here.

Heat map

What it needs

The intuition

When to use it

Where the estimator comes from

What it assumes

Statistical properties

How to read the output

A worked reading

Symptom → diagnosis → what to do

Common mistakes

Where to read this properly

See also

Related Charts

All analyses

Charts

Maps

Explore

Regression

Time series - one series

Time series - several series

Panel data

Multivariate statistics

Marketing research

Causal

Finance & risk

Actuarial

Tests & breaks

Machine learning

Econometrics that explains itself.