Understanding Causality: Philosophical Foundations, Causal Networks, and Simpson’s Paradox

In everyday life causality is ubiquitous, yet defining it rigorously is challenging; the article begins by asking readers to formulate a personal definition and then presents Hume’s view that constant conjunction implies causation, while highlighting the need for an interventionist perspective.

It explains the INUS condition (Insufficient but Necessary parts of an Unnecessary but Sufficient condition) with examples such as lightning, dry hay, and firefighters in fire incidents, illustrating why some sufficient conditions are not true causes.

The piece critiques Granger causality for capturing only predictive association without intervention, and briefly lists econometric tools (instrumental variables, regression discontinuity, difference‑in‑differences, propensity scoring, synthetic control) used for causal inference.

Epidemiology’s role is illustrated by John Snow’s cholera investigation, emphasizing randomized controlled trials, cohort studies, and case‑control designs as methods to uncover causal links in health research.

Three basic causal‑network configurations derived from Bayesian networks are described:

Chain (A → B → C) : controlling the mediator B can block the flow from A to C.

Fork (A ← B → C) : B is a common cause; adjusting for B removes spurious association between A and C.

Collider (A → B ← C) : conditioning on B opens a path between A and C, potentially creating a false association.

The article then applies causal networks to the Simpson’s paradox, using a fictional drug D study where aggregated data suggest the drug is beneficial, but stratified by gender the drug increases heart‑attack risk for both men and women. By recognizing gender as a confounder (a fork structure) and controlling for it, the correct conclusion is that drug D harms all groups.

A second example with drug B shows a different network where blood pressure acts as a mediator (chain); since the analysis does not aim to separate direct and indirect effects, no adjustment is needed, leading to the conclusion that drug B is beneficial.

Finally, the article presents visual examples of Simpson’s paradox in exercise‑cholesterol data, prompting readers to consider how causal reasoning determines whether exercise lowers cholesterol for specific subpopulations.

References include Judea Pearl’s “The Book of Why” and various online resources on causal inference and epidemiology.