Interactive statistics

56 free interactive lessons: from mean and median to A/B tests and Bayes. Drag the charts and build intuition. No sign-up.

Module 3: Distributions

How to tell which distribution your data has

We covered a family of distributions. But in real work nobody labels the data "this is log-normal". You have to recognize the shape yourself — it decides which measures and methods are even allowed. Let's assemble a practical checklist for "reading" a distribution.

PDF — density: the height itself is not a probability; the probability of an interval = the area under the curve over it. Total area = 1.

value →
▏ theory (PDF)▮ sample (histogram)Mean ≈ 0.00 · values drawn: 0

A symmetric "bell": most values crowd near the center, thinning toward the edges. μ sets where the center stands, σ how widely values fan out (σ stretches the bell wider and lower, μ slides it along the axis). Appears wherever a quantity is the sum of many small independent random influences — heights, measurement errors, part deviations. Press "pour" — the real sample\'s histogram approaches the theoretical density.

Step one — always draw a histogram (and a box plot). One glance answers the essentials: one hump or several? Symmetric or tailed? Any outliers? Train your eye right here: switch shapes in the widget's list (uniform, normal, exponential, log-normal, power law) and press "pour" each time — the yellow histogram of a real sample settles onto the theoretical curve, and each shape's silhouette sticks in memory.

What it means

An analyst's reflex: opened a new column — drew a histogram and a box plot, and only then computed anything. It catches skew, outliers, bimodality and tells you whether the mean can be trusted at all.

You rarely need the distribution's exact name; what matters is answering three questions — is it symmetric, is the tail heavy, is it one group. The answers drive the choice of measures (mean vs median) and tests (t-test vs non-parametrics).

Where it shows up

Service response time almost always has a heavy tail — which is why engineers go straight to percentiles, not the mean. Recognized the shape → picked the right metric.

Bimodal distributions are everywhere in product data: "time to purchase" is often two-humped (impulse vs deliberate buys), and analyzing them as one whole is a mistake.

Definitions
Bimodality
two humps in a distribution — almost always a sign of a mixture of two different groups; a cue to slice the data.
Q-Q plot
comparing the data's quantiles with a reference (normal): a straight line — shapes match; bends at the ends — heavy tails.
Poisson signature
in count data, variance ≈ mean — a hint that this is a stream of rare events.
Log transform
the logarithm compresses a long right tail; if a bell emerges afterward, the original was log-normal.
Next →
Enjoying the materials?

This site is built by the DataSlice Telegram channel: data analytics in plain words — real cases, metrics, careers. The channel is in Russian.

Open the Telegram channel