AI summary by TwinMind
Allen Downey | Probably Overthinking It | Talks at Google
Talks at Google
48m 58s
Introduction and Background
- Alan Downey introduced at Google as prolific author, professor, former Google employee (2009-2010)
- Worked on network performance analysis at Google, developed statistics library that evolved into empirical disc Python package
- Latest book "Probably Overthinking It" originated from blog articles starting with relay race observation in 2011
- Chapter 7 focus: "Causation Collision and Confusion" - serious consequences for lives
Low Birth Weight Paradox (1971)
- California researcher studied relationship between maternal smoking and health outcomes
- Key findings:
- Babies of smokers 6% lighter, more likely classified as low birth weight
- Low birth weight babies had higher mortality: 170 per thousand vs 8 per thousand background rate
- Paradox: Among low birth weight babies only, babies of smokers had ~50% lower mortality than non-smokers
- Media response: "mothers needn't worry smoking little risk to baby"
- 2014 retrospective showed this held up anti-smoking measures for ~decade
- Real-world consequences for public health policy
Obesity Paradox (1999)
- Study examined obesity, kidney disease, and mortality relationship
- Established facts: obesity causes kidney disease, kidney disease increases mortality, obese people have higher general mortality
- Paradox: Among kidney disease patients, obese patients lived longer
- Speculation: overweight patients have more adipose tissue, less likely to suffer energy deficits
- Led to decades of confusion (2006-2019 papers): "fact or fiction," "still a controversy," "perception versus knowledge"
Twin Paradox (2000)
- Relationship between multiple births (twins/triplets) and mortality by gestational age
- Known facts: multiple births more likely pre-term, pre-term births have higher mortality, multiple births have higher overall mortality
- Paradox: Among pre-term births only, multiple births had lower mortality than single births
- Proposed mechanism: twins in better health prior to 36-37 weeks gestation but don't benefit as much from additional time
Statistical Explanation: Berkson's Paradox
- All three paradoxes are examples of collision bias (form of sampling bias)
- Demonstrated with SAT scores example:
- General population: math and verbal scores correlated at r=0.7
- Elite University (total score ≥1320): correlation becomes r=-0.3 (anti-correlated)
- Safety College (1000≤total≤1300): correlation becomes r=-0.8 (strongly anti-correlated)
- Correlation varies from +0.7 to -0.8 depending on selection criteria
Real-World Examples of Berkson's Paradox
- Dating: "triangle of acceptable partners" - handsome men show full range nice to mean, ugly men must be nice to be acceptable
- Movies vs books: only hear about combinations where one is good (bad movie + bad book never gets attention)
- Restaurant quality: successful restaurants either have good location OR good food - if not one, more likely the other
Causal Diagrams and Collision Bias
- Two ways to enter sample creates the bias
- Low birth weight has two causes: maternal smoking (less harmful) + birth defects/other factors (more harmful)
- "Burnt toast" analogy: relieved when smoke in kitchen is just burnt toast vs house fire
- Doctor relieved when low birth weight explained by smoking vs birth defects
2006 Resolution Study
- Hernandez-Diaz, Schisterman, and Hernan explained low birth weight paradox
- Used 3 million babies from National Center for Health Statistics (1991)
- Replicated original findings, showed crossover point at ~2000g birth weight
- Above threshold: higher mortality for smokers' babies (expected)
- Below threshold: lower mortality for smokers' babies (Berkson's paradox)
Causal Diagram Terminology
- Collider: two arrows coming into single effect (creates bias when used for sampling/control)
- Mediator: A→B→C, B mediates relationship
- Confounder: single cause affecting both variables
- Rules: controlling for confounders good, controlling for colliders bad, controlling for mediators removes causal effect
Contemporary Examples and Warnings
- Recent studies (last year): diabetics eating ice cream associated with lower heart problems, overweight people + dairy desserts associated with reduced insulin resistance
- Research publication bias: two ways to get attention (novel OR good) - more novel = less likely correct
- Media negativity bias: negative stories more likely published and noticed
- Customer sampling bias: two ways to become customer (like product OR no choice) - affects satisfaction correlations
Practical Applications
- Restaurant advice: order least appetizing menu item (survived because surprisingly good vs sounds good)
- Tyler Cowen's observation: dishes survive on menu two ways (sounds good OR surprisingly good)
- "Perfect day" satirical advice: find unattractive partner (they'll be nice), dingy restaurant (food will be good), movie based on bad book
Q&A Session Highlights
- Selection geometry: different admission criteria create different correlation patterns
- Confirmation bias vs sampling bias distinction
- Terms useful as vocabulary and checklist, not rigid categories
- Causal diagrams help determine appropriate statistical controls
- Balance between healthy skepticism and rejecting all scientific knowledge
- Vaccines example: despite potential biases, strong evidence vaccines prevent illness and death
