Do-calculus

As explained in Causal AI

Do-calculus is a formal system of rules for transforming causal and interventional probability expressions using graphical-model assumptions.

When a model must answer what would happen if a variable were deliberately set, correlation is not enough. Do-calculus transforms intervention-based probability expressions into quantities that can be estimated from observational data, when a causal graph supplies the required assumptions. section 1.5.3 section 10.4.1

The system contains three if-then rules. Each rule links a graphical condition to an equivalent probability query: if variables are d-separated after graph surgery on a DAG, a probability expression may be inserted, removed, or transformed. The rules derive from d-separation, ideal interventions, and probability laws rather than intuition alone. section 10.4.1

A practical workflow treats do-calculus as an identification step between causal modeling and statistical estimation:

• Represent the causal question as an intervention, such as the target distribution under setting X to x. section 10.4.1
• Use graph structure and d-separation to transform the intervention expression into an observational estimand. section 10.4.1
• Estimate the resulting terms with statistical or machine-learning models, while weighing bias, variance, scalability, and parallelizability. section 1.5.3
• Automate identification with graphical software, or use a known result such as the backdoor adjustment estimand. section 10.4.1 section 10.4.2

Writer type and image content

To estimate the causal effect of writer type on images when a randomized assignment is infeasible, a DAG and do-calculus can derive an expression for conditional average treatment effects. The left side represents hypothetical interventions; the right side contains terms estimable from labeled image data. A related derivation applies rules 2 and 3 to obtain a backdoor estimand, with c-component factorization supplying an intermediate step. section 1.5.3 section 10.4.2

Figure 1.8 Visualization of an example CATE of writer type on an image. It is the pixel-by-pixel difference of the expected image under one intervention (W=_{"high school"}) minus the expected image under another intervention (W=_"bureaucrat"), with both expectations conditional on being images of the digit 4 with a certain level of thickness.

Compare two interventions as a pixel-by-pixel difference between their conditional expected images, making the causal query concrete. section 1.5.3

How Causal AI covers it

How deep it goes

Derives rules and a backdoor estimand from graph surgery, while emphasizing intuition and automation over memorizing every rule.

Examples

Uses writer-type effects on handwritten images and an alternative do-calculus derivation of the backdoor estimand.

What it assumes

Assumes causal graphs and probability basics, while building intuition for d-separation, interventions, identification, and causal queries.

What you can do after

Apply it via graphical identification libraries or derive estimands such as backdoor adjustment from a causal graph.

How it connects

Connects d-separation, ideal interventions, directed acyclic graphs, c-component factorization, and backdoor adjustment.

liveBookmanning.com

Read Causal AI on liveBook Get the book at manning.com

Related concepts

• causal identificationused by · Causal identification is the process of determining whether a causal quantity can be uniquely computed from available observational or interventional information and model assumptions.
• directed acyclic graphuses · A directed acyclic graph is a graph with directed edges and no directed cycles.
• causal adjustmentalternative to · Causal adjustment estimates a causal effect by conditioning on variables that block noncausal pathways and averaging over their target-population distribution.

This explainer was generated by Manning's knowledge graph from the text of Causal AI. The section links point to the parts of the book it draws on.