ECDF and Rug Plots

A histogram has to make a choice — where to put the bin edges — and that choice can quietly reshape the story. What if you could see a distribution without choosing anything: no bins to set, no smoothing bandwidth to tune, nothing averaged away?

That is the appeal of two underused views. The ECDF plots the running proportion of your data, and the rug plots the raw observations themselves. Neither hides anything behind a parameter, which makes them honest companions to the histogram.

The ECDF: a running tally of proportions

The empirical cumulative distribution function (ECDF) answers a precise question at every point on the x-axis: what fraction of the observations are less than or equal to this value? Sort the data, then walk left to right; each observation steps the curve up a little, so it climbs from 0 at the smallest value to 1 at the largest.

There are no bars and no curve-fitting here — just the data, accumulated. Every observation is represented exactly once, in its true position.

How to read it

The y-axis is a proportion, from 0 to 1, and that makes the curve a percentile-reading machine:

• The median is wherever the curve crosses y = 0.5 — drop straight down to the x-axis to read it.
• Any percentile works the same way: the 90th percentile sits where the curve reaches y = 0.9, the 25th where it reaches 0.25, and so on.
• Steepness encodes density. Where the curve rises steeply, many observations are packed into a small x-range — that's a dense region (a peak in the histogram). Where it is flat, data is sparse (a valley or a gap).

So the same modality a histogram shows as peaks appears in an ECDF as steep stretches separated by flatter ones — readable, though less pictorial. In exchange you get exact percentiles straight off the axis, with no binning required.

QuestionSelect one

On an ECDF of flipper_length_mm, you find the x-value where the curve crosses y = 0.5. What have you just read off?

The mean flipper length.

The median flipper length — the value with half the observations at or below it.

The most common flipper length.

The total number of penguins.

Its superpower: comparing groups

A histogram struggles to compare more than a couple of groups, because overlapping bars occlude one another. ECDF curves have no such problem: each group is a single thin line, and lines can cross and overlap without hiding each other. That makes the ECDF one of the best tools for comparing several distributions at once. Just map a categorical column to hue:

Read it directly: at any flipper length, the curve that is higher has a larger share of its penguins at or below that length — that is, it sits further to the left overall. Here Gentoo's curve climbs last and furthest right, confirming it as the long-flipper group, while Adelie and Chinstrap rise earlier on the left. Three distributions compared cleanly, no occlusion, no bins.

The tradeoff

Nothing is free. The ECDF's great virtue — it hides nothing — comes with a cost in intuitiveness:

• What you gain: no bin edges to bias the picture, no smoothing bandwidth to choose. Percentiles, medians, and group comparisons are exact and uncluttered. Nothing is hidden or invented by a parameter.
• What you give up: humans read shape and modality far more naturally from a histogram's literal humps than from the slope changes of a cumulative curve. A two-peaked distribution is obvious as two bumps; as two steep stretches in an ECDF it takes a trained eye.

So the two views are partners, not rivals. Use a histogram when you want someone to see the shape at a glance; use an ECDF when you need exact percentiles or a clean comparison of several groups. Showing both is often the most honest move of all.

Rug plots: the raw data, one tick at a time

A rug plot is the most literal distribution view there is: it draws a short tick mark for every single observation along an axis, like fringe on a rug. On its own it is spare, but as an add-on it is invaluable — laid under a histogram, a KDE, or beside a scatter, it shows exactly where the real data points sit, with no binning or smoothing in between.

Here it is on its own, via the axes-level sns.rugplot:

Its real value shows when you layer it onto another plot. displot will add one for you with rug=True, so the smooth or binned summary sits above the actual observations that produced it:

The rug is just as natural as fringe along the axes of a scatter plot, where it shows how each variable is distributed in one dimension while the dots show the joint relationship:

Your turn

Check your understanding

QuestionSelect one

What does the height of an ECDF curve at a given x-value represent?

The count of observations equal to that x-value.

The proportion of observations less than or equal to that x-value.

The probability density at that x-value.

The number of bins covering that x-value.

QuestionSelect one

Why is an ECDF often a better choice than overlapping histograms for comparing several groups on one axis?

ECDFs use more bins, so they capture finer detail.

Each group is a single line, so multiple curves can cross and overlap without hiding one another.

ECDFs show the mode of each group more clearly than a histogram.

ECDFs automatically remove outliers.

QuestionSelect one

When is a rug plot the right tool?

As the main view for a column with tens of thousands of observations.

As an add-on to another plot, showing the raw individual positions when points are sparse enough to distinguish.

To display the relationship between two numeric variables.

To compute and display group means.

You now have two bin-free, low-distortion ways to read a distribution: the ECDF for exact percentiles and clean group comparisons, and the rug for the raw observations themselves. Together with the histogram and KDE, they give you a full toolkit for seeing a single variable from every angle. Next we leave one-variable distributions behind and turn to categorical plots.