What is causal discovery?
Causal discovery finds which variables in a table drive which others. What it is, how it differs from correlation and prediction, what it needs from your data, and where its limits are.
In short
Causal discovery is a family of methods that read a table of observations and propose a causal graph: a map of which variables directly influence which others, and in which direction.
A correlation says that two numbers move together. A causal graph says which one moves the other, along which path, and therefore what is likely to happen when you change one of them yourself.
Why correlation is not enough
Most business and research data is observational: it was recorded as things happened, not produced by a controlled experiment. In such data, many variables move together without one causing the other.
Take a delivery business. Late deliveries and overtime hours rise together. Overtime does not cause the delays: both rise on busy days when the team is short-staffed. Cutting overtime would not reduce delays; adding staff on peak days might. A dashboard shows the two lines rising together; a causal graph shows the common cause behind them.
Prediction has the same blind spot. A predictive model answers "what will this number be, given what I see?" It may lean on a variable that only moves with the outcome. Changing that variable on purpose then changes nothing. Decisions need the other question: "what happens if I act?"
How it works
Causal discovery methods look for the traces that cause and effect leave in data. There are three main kinds of clue:
- Independence patterns. If A affects C only through B, then A and C stop being related once B is held fixed. Constraint-based methods, such as PC and FCI, test these patterns column by column and build the graph from them.
- Scores. Score-based methods, such as GES, BOSS and GRaSP, search through possible graphs and keep the one that explains the data well with as few links as possible.
- Asymmetries. When relationships are not purely linear with bell-shaped noise, a cause and its effect leave different footprints. Methods such as LiNGAM use the shape of that noise to tell which way a link points.
The result is a graph. Each node is a column of your table; each arrow is a proposed direct causal link. When the data cannot tell the direction of a link apart, a method may leave it undirected rather than guess.
Discovery, effect estimation and what-if
Causal discovery is the first of three questions that causal AI answers:
| Step | The question | What you get |
|---|---|---|
| Causal discovery | What causes what? | A causal graph |
| Effect estimation | How much does X change Y? | The size of an effect |
| Counterfactual analysis | What would Y have been if X had been different? | An estimated outcome for a scenario |
The graph comes first because the other two depend on it: to estimate the effect of X on Y, you need to know which other variables to account for, and the graph tells you.
What it needs from your data
- One table, one row per observation (an order, a day, a patient, a sample) and one column per variable, holding numbers.
- Enough rows. More rows let weaker links be found and make the result steadier; with few rows, only strong links stand out.
- The variables that matter. A common cause that is missing from the table can make two other variables look causally linked. Some methods, such as FCI, are designed to flag where such a hidden cause may be at work.
- No identifiers or copies. Row numbers, IDs and columns that are copies of one another add noise, not information.
In UCMT Labs, a data check runs on your file before discovery. It points out ID-like columns and other problems, and suggests a starting method and a method to compare it with, with the reasons.
What it cannot do
- It proposes; it does not prove. Every method rests on assumptions about the data. When they do not hold, a link can be missing, extra or reversed.
- Some directions stay open. Several graphs can explain the same data equally well; the methods then leave the direction undecided.
- The result depends on the columns you include. Add or remove a variable and the graph can change.
- Important links deserve a second check: your knowledge of the domain, a comparison of several methods on the same data, or an experiment where one is possible.
Try it in UCMT Labs
- Create a free account and open causal inference. Start with a sample dataset or upload your own CSV file.
- Read the data check and the suggested methods.
- Run causal discovery, then compare several methods to see which links they agree on.
- Ask what-if questions on the graph in the Reasoning tab.
The user guide walks through each step with screenshots.