Concepts · no installation required
The ideas the rest of this site assumes
The tutorials show axiom doing things. These pages explain what the things are, for a reader who has not met them — what makes a question causal, why a confounder is not just another variable, and why a study that was right where it ran can be wrong where it is used.
You do not need Python to read any of them. Where a page shows code, the output beside it is what that code printed when this page was built.
If you are short of time
Three ways through
“I need to make a decision from data I already have.”
Read prediction is not intervention, then confounders, colliders and mediators, then identification comes before estimation. That is enough to know whether your data can answer your question at all — which is usually the finding.
“Somebody handed me a number and I have to use it.”
Go straight to the estimand in eight facets and then why a model does not travel. Most numbers that turn out to be wrong were not wrongly computed. They were correctly computed for a slightly different question.
“I have to defend this in six months.”
Provenance first, then what would have to be true to overturn this. The useful artefact is not the estimate. It is the list of things you assumed and what each one would take to break.
The shelf
Twelve guides, in order
Each stands on its own, and each assumes the ones above it. Fifteen minutes a page, give or take.
Shelf 1 · 4 guides
Foundations
What a causal question is, and why the usual tools do not answer one.
01
Prediction is not intervention
A model that forecasts an outcome perfectly can be silent about every decision you might take. Why, and what the difference is.
Read it
02
Confounders, colliders and mediators
The same variable, in three positions. Adjusting for the wrong one does not fail loudly — it returns a number.
Read it
03
Writing your beliefs down
A causal graph is not a diagram of the data. It is a set of claims about mechanism, and it is testable.
Read it
04
Identification comes before estimation
Whether an effect is recoverable at all is a question about the graph, not the sample. It is answerable before you fit anything.
Read it
Shelf 2 · 4 guides
Writing the question down so it travels
The part most workflows skip, and the reason a sound number ends up answering the wrong question somewhere else.
05
The estimand, in eight facets
Two quantities are the same quantity only if all eight match. Six of the eight can be bridged; two cannot.
Read it
06
Dimensions, units and scope
Three different kinds of mismatch that all look like a unit problem. Only one of them is.
Read it
07
Why a model does not travel
A selection diagram, four things it can say, and the formula that carries an effect from where it was measured to where it is used.
Read it
08
Provenance: hashes, ledgers, and the paper trail
Every number carries how it got here — or it is not evidence, it is a number in a slide.
Read it
Shelf 3 · 4 guides
Doing the work
Where the ideas above turn into decisions about what to measure and what to believe.
09
Priors that mean something
A prior is a claim about magnitudes in units you can argue about, not a regularization knob.
Read it
10
Deciding what to measure
Power is not the question. What the decision turns on is the question, and it changes what you should go and measure.
Read it
11
Folding an experiment in
An experiment measures one thing precisely. Calibration is how that one thing constrains a model of everything else.
Read it
12
What would have to be true to overturn this?
The most useful output of an analysis is usually the size of the confounder that would erase it.
Read it
Then what