Data Analytics: Statistics & Data Visualization: all lessons

P(positive | ill) and P(ill | positive) are very different numbers. Swapping them is the most expensive mistake in applied statistics.

Take the mean of many samples and their distribution approaches a normal curve — almost whatever the shape of the original data. It is why most inferential statistics works.

A p-value answers a very specific question — and almost never the one you want to ask. Power tells you whether your experiment could find anything in the first place.

What "95%" actually promises, and why an interval is more honest than a single p-value.

Hypotheses, primary metrics, guardrail metrics and the unit of randomisation — the decisions to make before a single visitor enters the experiment.

The minimum detectable effect decides how many visitors you need, and the calendar decides how long — always in full weeks.

A proportion test, a confidence interval for the difference, continuous guardrail metrics, and how to report the result to the people who have to decide.

Peeking, sample ratio mismatch, novelty effects and too many metrics — four habits that feel reasonable and make a tidy experiment give the wrong answer.

Slope, intercept and R² — and what R² does not tell you.

Why a coefficient changes when another variable is added — and when that change fixes the answer, and when it breaks it.

Residuals, heteroscedasticity and overly influential points — four checks that decide whether that tidy coefficient table can be trusted.

Identical means, standard deviations and correlations can hide data of completely different shapes. One chart catches what all those numbers miss.

An exploratory chart finds the result; a presentation chart delivers it. Titles that state the finding, direct labels, and removing everything that does not help.

Categorical, sequential and diverging palettes do three different jobs — and one in twelve male readers cannot tell your red from your green.

A Figure holds Axes, an Axes holds Artists — once you hold the objects, every element of a chart is controllable, and a three-panel figure stops being scary.

Distributions, group comparisons and facets in one line from a DataFrame — as long as the data is in long format and you know which chart hides what.

Hover, zoom and filters turn a chart from an answer into a tool for asking — useful for exploration and for sharing, and harmful on a projector.

A dashboard that gets used starts from the decision someone has to make every morning — not from the data that happens to exist.

A KPI is its definition. North-star, input and guardrail metrics — and how to spot a metric that is easy to game or that only moves because of noise.

Context, tension, finding, recommendation — the structure that gets people to act on a correct analysis instead of merely nodding.

Bars encode length, so bars must start at zero; lines encode position, so they need not. And two y-axes with free scales can make any two series look like they move together.

The same series can be "down 19%" or "up 10%" depending on where you start; a cumulative chart never falls; and an icon twice as tall looks four times as big.