The Book of Why: How a 'causal revolution' is shaking up science

A much-needed "causal revolution" has arrived in Judea Pearl's 'The Book of Why'. But despite vast improvements over "trad stats", there's cause for concern over logic-losing numbers.

Illustration by Julia Suits, author of The Extraordinary Catalog of Peculiar Inventions, and The New Yorker cartoonist.
Illustration by Julia Suits, author of The Extraordinary Catalog of Peculiar Inventions, and The New Yorker cartoonist.

1. The Book of Why brings a “new science” of causes. Judea Pearl’s causology graphically dispels deep-seated statistical confusion (but heterogeneity-hiding abstractions and logic-losing numbers lurk).


2. Pearl updates old correlation-isn’t-causation wisdom with “causal questions can never be answered from data alone.” Sorry, Big Data (and A.I.) fans: “No causes in, no causes out” (Nancy Cartwright).

3. Because many causal processes can produce the same data/stats, it’s evolutionarily fitting that “the bulk of human knowledge is organized around causal, not probabilistic relationships.” Crucially, Pearl grasps that “the grammar of probability [& stats]… is insufficient.”

4. But trad stats isn’t causal “model-free,” it implicitly imposes “causal salad” models—independent factors, jumbled, simple additive effects (widely method-and-tool presumed ... often utterly unrealistic).

5. “Causal revolution” methods enable richer logic than trad-stats syntax permits (for instance, arrowed-line causal structure diagrams enhance non-directional algebra).

6. Paradoxically, precise-seeming numbers can generate logic-fogging forces. The following reminders might counter rote-method-produced logic-losing numbers.

7. Causes of changes in X, need not be causes of X. That’s often obvious in known-causality cases (pills lowering cholesterol aren’t its cause) but routinely obfuscated in analysis-of-variance research. Correlating variation percentages to factor Y often doesn’t “explain” Y’s role (+see “red brake risk”). And stats factor choice can reverse effects (John Ioannidis).

8. Analysis-of-variance training encourages fallacy-of-division miscalculations. Many phenomena are emergently co-caused and resist meaningful decomposition. What % of car speed is “caused’ by engine or fuel? What % of drumming is “caused” by drum or drummer? What % of soup is “caused” by its recipe?

9. Akin to widespread statistical-significance misunderstandings, lax phrasing like “control for” and “held constant” spurs math-plausible but impossible-in-practice manipulations (~“rigor distoris”).

10. Many phenomena aren’t causally monolithic “natural kinds.” They evade classic causal-logic categories like “necessary and sufficient,” by exhibiting “unnecessary and sufficient” cause. They’re multi-etiology/route/recipe mixed bags (see Eiko Fried’s 10,377 paths to Major Depression).

11. Mixed types mean stats-scrambling risks: fruitless apples-to-oranges stats like average humans have 1 testicle + 1 ovary.

12. Pearl fears trad-stats-centric probability-intoxicated thinking hides its staticness, whereas cause-driven approaches illuminate changing scenarios. Causality always beats stats (which encode unnovel cases). Known causal-composition rules (your system’s syntax) make novel (stats-defying) cases solvable.

13. “Causal revolution” tools overcome severe trad-stats limits, but they retain rush-to-the-numbers risks (is everything relevant squeezable into path-coefficients?) and type-mixing abstractions (e.g., Pearl’s diagram lines treat them equivalently but causes work differently in physics versus social systems).

14. “Cause” is a suitcase concept, requiring a richer causal-role vocabulary. Recall Aristotle’s cause kinds—material, formal, proximate, ultimate. Their qualitative distinctness ensures quantitative incomparability. They resist squashing into a single number (ditto needed Aristotle-extending roles).  

15. Causal distance always counts. Intermediate-step unknowns mean iffier logic/numbers (e.g., genes typically exert many-causal-steps-removed highly co-causal effects).

16. Always ask: Is a single causal structure warranted? Or casual stability? Or close-enough causal closure? Are system components (roughly) mono-responsive?

17. Skilled practitioners respect their tools’ limits. A thinking-toolkit of context-matched rule-of-thumb maxims might counter rote-cranked-out methods and heterogeneity-hiding logic-losing numbers.

 

A 62-year old Russian mystery (and conspiracy theory) has been solved

Some mysteries take generations to unfold.

Winter in the Ural Mountains

Credit: Hикита Чертков / Adobe Stock
Surprising Science
  • In 1959, a group of nine Russian hikers was killed in an overnight incident in the Ural Mountains.
  • Conspiracies about their deaths have flourished ever since, including alien invasion, an irate Yeti, and angry tribesmen.
  • Researchers have finally confirmed that their deaths were due to a slab avalanche caused by intense winds.
Keep reading Show less

Could muons point to new physics?

New data have set the particle physics community abuzz.

Credit: Stefano Garau / Adobe Stock and Trahko / Adobe Stock
13-8
  • The first question ever asked in Western philosophy, "What's the world made of?" continues to inspire high energy physicists.
  • New experimental results probing the magnetic properties of the muon, a heavier cousin of the electron, seem to indicate that new particles of nature may exist, potentially shedding light on the mystery of dark matter.
  • The results are a celebration of the human spirit and our insatiable curiosity to understand the world and our place in it.
Keep reading Show less
Credit: William Thomas Cain via Getty Images
Personal Growth
  • Benjamin Franklin wrote essays on a whole range of subjects, but one of his finest was on how to be a nice, likable person.
  • Franklin lists a whole series of common errors people make while in the company of others, like over-talking or storytelling.
  • His simple recipe for being good company is to be genuinely interested in others and to accept them for who they are.
Keep reading Show less
Quantcast