Make your data sing.
Handed a new dataset, most of us group by a column, sum a number, sort, and stop. "The most common name is James." True, but nobody leans forward. Better questions come from noticing what kinds of columns you have and which ones appear together.
The whole framework in five notes. Each step climbs from a flat group-by toward a question worth answering. Click a note to jump to it.
Group by name, sum, sort. Done.
Data: U.S. Social Security Administration baby names by state, 1910–2021 (names with fewer than 5 babies in a state and year are left out). Hover the charts for values.
Name each column's job
Every column does one of five jobs. Say it in plain English. Bottom row: the five columns of the baby-names dataset, one per job.
💡 Some columns have two jobs. A year is a When, but also an ordering. A state is a Where, but also a Kind that nests inside a region. Spotting double jobs is how you find step 4.
Every pair of types asks its own question
Pick two columns and find their types on the grid. The cell is a ready-made question. Outlined cells are the richest pairs: start there.
Nine moves, three families
A move is a small transformation that turns a flat answer into an interesting one. Almost every good question uses at least one. Try to use moves from two different families.
Moves in action
Jennifer shot up and fell back within one generation. Elizabeth never left. Who × When + peak, measured as a share, so baby booms don't distort it.
The most common names in Utah look like everywhere else. The most distinctive ones don't. Who × Where + over-index.
The pattern breaks: a name switches sides. Who × Kind × When + share + reversal.
The name itself is a column
Treat the text as data: its length, first letter, last letter, number of vowels. Then any other move applies.
Seed: “boys' names ending in n.”
Full question: “What share of boys born each year got a name ending in ‘n’, and when did it start rising?”
Knock on three doors
So far every row stands alone. When certain columns appear together, rows start to relate to each other and become a journey, a basket, a network or a hierarchy. Ask three questions of every dataset. Each "yes" opens a new toolkit, and the technical step is usually small.
Grow a seed into a full question
Good questions start as short, vague seed phrases. You can't compute a seed. A full question spells out five parts.
Checklist for any new dataset
- Type every column: Who, Kind, When, Where, How much.
- Pair types: pick two outlined cells and write their questions.
- Move: apply moves from at least two families.
- Shape: knock on all three doors. Order? Pair? Nest?
- Sharpen three seeds into full questions, and write your expected answer first.