Collider

As explained in Causal Inference for Data Science

A collider is a variable that is causally influenced by two other variables and whose conditioning can induce or alter their association.

A collider is a variable with two arrows pointing into it: X → Z ← Y. X and Y are both causes of Z. Unlike a confounder or mediator, Z does not create a causal route from X to Y; in the three-node collider case, only their direct path is causal. section 6.1.2 section 7.1.4

The practical risk appears when an analysis selects, groups, or adjusts for the collider. Conditioning on Z can create a correlation between its parents, or change one already present. This can make an association inside a selected group look different from the association in the full population. section 7.1.4

Acceptance can reverse an association

GPA and SAT scores can be positively correlated overall, while restricting attention to students accepted by a university can make their correlation negative. Qualification is a collider because both GPA and SAT contribute to it; analyzing only accepted students conditions on that collider. section 7.1.4

Figure 7.11 Student university acceptance DAG, based on GPA and SAT results. Qualification is a collider. Conditioning on a collider can change the correlation between the collider’s parents.

Use the acceptance DAG to trace the two arrows into Qualification, then ask how restricting the data to accepted students changes the GPA–SAT comparison. section 7.1.4

In a treatment trial, Side Effects can be a collider when both Treatment and Recovery affect it. To estimate the modeled effect of Treatment on Recovery, use a regression of Recovery on Treatment rather than adding Side Effects as an adjustment variable. This differs from confounder adjustment, where including the common cause is required. section 6.1.2

The mechanism can be checked with synthetic data: generate independent X and Y, let both affect Z, then compare their overall correlation with correlations within Z = 1 and Z = 0 groups. Conditioning creates correlations that were absent before selection. section 7.1.4

When classifying a third variable Z for an X-to-Y causal question, distinguish these patterns:

• Confounder: Z causes both X and Y; adjust for Z to estimate the modeled effect. section 6.1.2
• Mediator: X causes Z, which causes Y; adjusting for Z targets a direct effect, while omitting it targets the total effect. section 6.1.2
• Collider: X and Y cause Z; do not add Z when estimating the modeled X-to-Y effect. section 6.1.2

How Causal Inference for Data Science covers it

How deep it goes

Defines collider DAGs, derives correlation and regression quantities, and demonstrates conditioning with synthetic data.

Examples

Uses treatment side effects, university acceptance from GPA and SAT, dice sums, and simulated binary variables.

What it assumes

Assumes basic variance, covariance, correlation, and linear-regression equations while introducing collider structure through DAGs.

What you can do after

Evaluate/understand when conditioning or selection can induce misleading associations between a collider’s parents.

How it connects

Connects colliders to DAGs, confounders, mediators, conditional independence, correlation, and linear regression.

liveBookmanning.com

Read Causal Inference for Data Science on liveBook Get the book at manning.com

Also explained in

• 

  liveBookmanning.com
  Causal AIA collider is a variable in a causal graph with directed paths entering it from at least two other variables, so those paths converge at that variable.

Compare how 2 books explain collider →

Related concepts

• conditioningused by · Conditioning is restricting or guiding a probabilistic or generative process using specified information, variables, values, or events.
• collider biasproduces · Collider bias is bias caused by conditioning on a collider or its descendant, thereby inducing an association between its causes.
• d-separationcontrasts with · D-separation is a graphical criterion for determining whether a conditioning set blocks all relevant paths between variables in a graph, implying conditional independence under the graph’s assumptions.
• correlationcontrasts with · Correlation is a statistical relationship describing how two or more variables or quantities vary together, including the strength and direction of their association, without by itself implying causation.
• causal discoveryaddresses · Causal discovery is the process of inferring causal relationships or a causal graph from data under assumptions about the data-generating process.

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