STRATA
Skip to study
XLVII

V.09 / DISCOVERY

Causal inference and intervention

Under what assumptions can an observational estimate predict the effect of an intervention?

A source-led introduction1 source references1 related experiments

Start with the question.

Causal inference asks what would change under an intervention, which requires more structure than observing an association. Pearl's account makes assumptions about causal relationships explicit through structural models. A diagram is useful when it identifies confounding, mediation, selection, or a missing variable that changes what can be inferred. The next step is to connect that structure to a concrete intervention or identification strategy.

Explain a misleading association.

In an illustrative dataset, umbrella use and wet streets occur together. Rain can cause both, so removing umbrellas would not ordinarily dry the street. Draw those arrows, then contrast observing an umbrella with intervening to provide one. The difference clarifies why prediction from a variable and the effect of manipulating it answer different questions.

What would you inspect?

  • The intervention being estimated and the variables assumed to cause both exposure and outcome.
  • Adjustment choices, selection mechanisms, and unmeasured common causes.

Take the next step.

Draw a causal model for one public claim and state the data or intervention needed to identify the specific effect being asked about.

Put part of the question to work.

These models isolate a mechanism you can vary. Their assumptions define the part of this topic they can illuminate.

Try connectivity and removal.

Change one parameter, compare the numerical output, and export the result. The experiment loads when you open it.

Open its full page

Watch the quantities change.

12 seconds · Original model visualization · Silent

Connectivity and removal

Compare local connections with the paths that hold a network together.

Read the explanation and method

Forty-eight points are drawn from one fixed random seed. A link joins two points whenever their distance is within the displayed radius.

The radius grows from 0.05 to 0.4 of the unit square. Existing links remain while new links merge previously separate connected groups.

Orange marks the largest connected group. Connectivity here describes paths through this constructed graph; it does not measure information quality, influence, or causality.

Method. Two nodes share an undirected edge when their Euclidean distance is at most the connection radius. After removal, a depth-first traversal assigns every active node to a connected component. Groups are ranked by size; ties use their lowest node identifier.

Read the supporting material.

  1. Causal inference in statistics: An overview

    Judea Pearl / UCLA · Structural causal models, interventions, counterfactuals, and identification assumptions.

Follow the comparisons.

Read the neighboring questions.