Histograms

A scatter plot asks how two variables relate. A histogram asks a simpler, more fundamental question about one: where do its values pile up? Hand Seaborn a single numeric column and a histogram turns it into a shape — a silhouette you can read for peaks, spread, and stragglers in about a second.

The idea is humble. Chop the number line into equal-width bins, then count how many observations fall in each one. Tall bar, lots of values there; short bar, few. That count-by-bin picture is the most common way to see a distribution.

What a histogram needs and shows

• Data it needs: one numeric (continuous) variable. Optionally a categorical column for hue to compare groups.
• What it highlights best: the shape of a distribution — where it peaks (modality), whether it leans left or right (skew), how wide it is (spread), and whether stray values sit far out (outliers).
• What it hides or distorts: the bars depend entirely on where the bin edges fall. A different bin width can manufacture a bump or erase one, so a single histogram is one opinion about the data, not the truth.
• When it breaks: comparing many groups at once — overlapping colored histograms occlude each other into mud. We'll handle that with multiple= and faceting below.

The minimum

displot ("distribution plot") is Seaborn's figure-level entry point for distributions. With kind="hist" you get a histogram of whatever column you put on x:

That silhouette already says something: flipper lengths are not a single tidy hump. There is a tall group on the left and a second rise on the right — a hint that more than one kind of penguin is hiding in this column. Hold onto that observation; it is about to be the whole lesson.

Bin width is a choice, not a fact

Here is the single most important thing to understand about histograms: the number of bins (or equivalently their width) is a decision you make, and it changes the story the chart tells. Same data, same column — different number of bars, different conclusion.

Watch the same flipper lengths drawn three ways. First, far too few bins:

With only five fat bins the two groups blur together into one vague mound. The structure we glimpsed earlier — two populations — has been erased by coarse binning. Now swing to the other extreme:

Now there are so many skinny bars that each holds only a handful of penguins. The picture is jagged and noisy — random gaps and spikes that say more about sampling luck than about the real distribution. Somewhere in between sits a width that shows the genuine structure:

With about twenty bins the two peaks stand out clearly — the flipper lengths are bimodal, exactly the signature of two kinds of penguin mixed together. Too few bins hid it; too many buried it in noise; a sensible width let it speak.

QuestionSelect one

You draw a histogram of one variable with bins=5 and see a single smooth hump. A colleague redraws it with bins=20 and now sees two clear peaks. What is the most reasonable conclusion?

The bins=20 version invented peaks that aren't really in the data.

Five bins were too coarse and merged real structure; the finer view reveals genuine bimodality.

Histograms are unreliable, so neither picture means anything.

The data must have changed between the two plots.

bins vs binwidth

You have two ways to control resolution, and they answer different questions:

• bins=20 fixes the number of bars; their width is computed from the data range. Handy when you want a quick "more or fewer."
• binwidth=2 fixes the width of each bar in the variable's own units (here, 2 mm). Handy when the width is meaningful — "group ages into 5-year bands," "bill lengths every 2 mm" — and when you want two charts to share an identical scale.

If you set neither, Seaborn picks a reasonable width for you. That default is a fine starting point — but now you know to nudge it and see whether the shape holds.

A smooth overlay with kde=True

Because bin edges are arbitrary, you may want a smoother summary of the same shape. Adding kde=True overlays a kernel density estimate — a smooth curve fitted to the data — on top of the bars:

The curve traces the same two-humped shape without committing to hard bin edges. It has its own tuning knob (a smoothing bandwidth) and its own ways of misleading — so the KDE gets a full treatment on the next page. For now, just know it as a quick way to sketch the distribution's outline over the bars.

Comparing groups with hue

A histogram earns its keep when you split one column by a category and ask how the groups differ. Map a categorical column to hue:

The mystery is solved in color: the left peak is mostly Adelie and Chinstrap, the right peak is Gentoo. The bimodality was a mix of species all along — the same reveal a hue gave us on the scatter page, now for a single variable.

Choosing how the groups stack: multiple=

When colored histograms share an x-axis they have to coexist somehow, and the multiple= argument decides the arrangement. The four options suit different goals:

• "layer" (default) — bars are drawn semi-transparently on top of each other. Good for 2–3 groups that don't overlap much; beyond that the colors muddy together.
• "dodge" — bars for each group sit side by side within each bin. Makes per-bin comparison exact, but gets cramped fast with many groups or many bins.
• "stack" — groups are piled vertically. The total height shows the combined count, but only the bottom group sits on a flat baseline, so the upper ones are harder to compare.
• "fill" — like stack but every bar is normalized to the same full height, so you read each group's proportion within a bin rather than its count.

Another readable choice for overlapping groups is element="step", which draws each distribution as an outline instead of filled bars — the lines cross cleanly without hiding one another:

Counts or density? stat=

By default each bar's height is a raw count of observations (stat="count"). That is intuitive, but it makes groups of unequal size hard to compare — a bigger group has taller bars everywhere simply because it has more penguins. Switching stat rescales the y-axis:

• stat="count" (default) — number of observations per bin.
• stat="density" — bars scaled so the total area is 1, which is what lets a histogram sit on the same scale as a KDE.
• stat="probability" — bar heights sum to 1, i.e. each bar is the fraction of observations in that bin.

When you compare groups of different sizes, stat="density" (or "probability") is usually the fairer choice: it asks "what is each group's shape?" rather than "which group is bigger?".

A common mistake

The classic histogram error is reading one bin width as gospel. A report that shows a single smooth hump — drawn, unknowingly, with too few bins — can hide a second population entirely, and a decision gets made on a shape that was never really there. The fix is cheap: vary the width, and trust only the features that survive.

Your turn

Check your understanding

QuestionSelect one

What does a histogram fundamentally display?

The relationship between two numeric variables.

The distribution of one numeric variable, by binning its values and counting how many fall in each bin.

A ranking of categories from largest to smallest.

The proportion of a whole made up by each category.

QuestionSelect one

What is the practical difference between bins=20 and binwidth=2 in sns.displot(..., kind="hist")?

They are interchangeable aliases for the same setting.

bins=20 fixes the number of bars; binwidth=2 fixes each bar's width in the data's units.

bins works only for integers, binwidth only for decimals.

binwidth changes the data, while bins only changes the display.

QuestionSelect one

You want to compare the flipper_length_mm distributions of five penguin-like groups on one axis. A layered (overlapping) histogram is unreadable. What is the better move?

Keep layering but pick brighter colors.

Map all five groups to size instead of hue.

Use element="step" outlines, or give each group its own panel with col=.

Reduce the number of bins to 2.

QuestionSelect one

You are comparing two groups of very different sizes and want to compare their shapes, not which group has more observations. Which stat is the most appropriate?

stat="count", the default.

stat="density" (or "probability"), which normalizes each group.

stat="count" with more bins.

There is no way to compare groups of different sizes.

You can now turn a column of numbers into a shape and read it honestly, mindful that the bin width is a choice you control. Next we follow the smooth curve we glimpsed with kde=True and give the kernel density estimate the full treatment it deserves.