This summary of The Book of Why: The New Science of Cause and Effect by Judea Pearl and Dana McKenzie explains how causal models can help us to better understand and control the world around us. This was a difficult one and I'm not sure how intelligible the key takeaways alone will be, as the diagrams are crucial to this book.

Nevertheless, as usual I've set them out below and you can find the full summary (with diagrams) by clicking the link above.

KEY TAKEAWAYS

• ​Statisticians woefully misunderstood causation for a long time. In the 1980s, Pearl sought to better understand causality in the hopes of developing Artificial Intelligence that could match or exceed human-level intelligence.
• The Ladder of Causation is a metaphor to describe the different ways in which humans, animals and computers understand causation. There are 3 rungs:
  1. Seeing (observing correlations). Most animals and computers (as at 2017) were on this first rung.
  2. Doing (making interventions). Early humans and some animals are on this second rung.
  3. Imagining (counterfactual reasoning). Humans are on this third rung. Our ability to imagine what could have happened if one thing changed likely played a big role helping us understand and control our environment.
• Seeing has limits. Causation has a direction but correlations and probabilities do not. Data alone does not tell you anything about the underlying causal relationships.
• While seeing is passive, Doing is active. We understand how the world works by forming hypotheses, testing them with interventions, and then getting feedback on our actions. The ideal intervention is a randomised controlled trial (RCT), but this is not always feasible.
• Causal models can offer an alternative way to move from the first to the second or even third rungs on the Ladder of Causation. They help us:
  • condition on the right things (confounders) and not the wrong things (mediators and colliders); and
  • use data to reach causal conclusions based on certain assumptions.
• Limits of causal models.
  • The usefulness of your model depends on how accurately your diagram matches reality.
  • You can falsify models but you cannot prove a model is correct.
  • Some hypotheses can only be tested with interventions.

You can find the full detailed summary on the website. If you found this summary useful, consider forwarding to a friend you think might enjoy it.

Thanks for subscribing! Until next time,

​To Summarise​